326
Int. J. Data Mining and Bioinformatics, Vol. 8, No. 3, 2013
Template-based scoring functions for visualising biological insights of H-2Kb–peptide–TCR complexes I-Hsin Liu and Yu-Shu Lo Institute of Bioinformatics and Systems Biology, National Chiao Tung University, Hsinchu 30050, Taiwan E-mail:
[email protected] E-mail:
[email protected]
Jinn-Moon Yang* Core Facility for Structural Bioinformatics, Institute of Bioinformatics and Systems Biology, Department of Biological Science and Technology, National Chiao Tung University, Hsinchu 30050, Taiwan E-mail:
[email protected] *Corresponding author Abstract: Major Histocompatibility Complex (MHC), peptide and T-Cell Receptor (TCR) play an essential role of adaptive immune responses. Many prediction servers are available for identification of peptides that bind to MHC class I molecules but often lack detailed interacting residues for analysing MHC–peptide–TCR interaction mechanisms. This study considers both the interface similarity and the interacting force for identifying binding models. Our model, considering both the MHC–peptide and the peptide–TCR interfaces, is able to provide visualisation and the biological insights of binding models. We believe that our model is useful for the development of peptide-based vaccines. Keywords: MHC; major histocompatibility complex; TCR; T-cell receptor; template-based scoring function; peptide. Reference to this paper should be made as follows: Liu, I.H., Lo, Y.S. and Yang, J.M. (2013) ‘Template-based scoring functions for visualising biological insights of H-2Kb–peptide–TCR complexes’ Int. J. Data Mining and Bioinformatics, Vol. 8, No. 3, pp.326–337. Biographical notes: I-Hsin Liu is a PhD Student in the Institute of Bioinformatics and Systems Biology, National Chiao Tung University, Taiwan. Her major research interests are related to protein–ligand interaction and proceeding with the peptide vaccine. Yu-Shu Lo is a PhD Student at the Institute of Bioinformatics and Systems Biology, National Chiao Tung University. His research interests are in the areas of protein–protein interaction and protein interaction family.
Copyright © 2013 Inderscience Enterprises Ltd.
Template-based scoring functions for visualising biological insights
327
Jinn-Moon Yang is a Full Professor and Director of the Institute of Bioinformatics and Systems Biology, National Chiao Tung University, Taiwan. His research focuses on computer-aided drug discovery, structural bioinformatics and systems biology. His h-index is 16 and has published over 40 SCI papers on some journals, such as Genome Biology, Nucleic Acids Research and Bioinformatics. His team has developed a molecular docking tool, namely GEMDOCK, which is one of the widely used docking tools in the world. For structural bioinformatics, his team achieved successful results on a fast protein structure search tool (3D-BLAST), which is as fast as BLAST and has the features of BLAST. Finally, his team has proposed protein–protein interacting family, which is similar to the concept of the protein or ligand family.
1
Introduction
Peptides are important biopolymers that are involved in most cellular processes (Lin et al., 2008), especially in adaptive immune responses. TCRs recognise short peptides, the processing products of protein antigens, in association with class-I MHC molecules to induce the immune responses. Therefore, identification of peptides that will be processed and presented with MHC molecules is a great benefit for developing of peptide-based vaccines. Many computational methods have been developed to predict MHC–peptide interactions by using SVM (Donnes and Kohlbacher 2006), known three-dimensional (3D) structure-based modelling (Altuvia and Margalit, 2004; Jojic et al., 2006) and matrix-based methods (Rammensee et al., 1999; Hakenberg et al., 2003). However, these methods provide highly predicting accuracy and the detailed binding models are insufficient for the mechanisms analysis. A known 3D structure of binding interface provides interacting residues and atomic details for thousands of direct physical interactions. According to our present knowledge, it is usually useful to build an interaction model of two proteins by comparative modelling if a known complex structure comprising homologues of these two sequences is available (Aloy and Russell, 2002; Lu et al., 2003). Since with the increase in MHC–peptide–TCR crystal structures, the concept from protein–protein interaction can be applied to the MHC–peptide and peptide–TCR interactions. To address these issues, we numerously enhanced and modified our previous study, 3D-domain interologues with template-based scoring function (3D-template PPI prediction method (Chen et al., 2007)). We evaluated the idea on the mouse class-I MHC (H-2Kb)–peptide and peptide–TCR interaction. Our scoring function has the similar predicting accuracy of the H-2Kb–peptide interface with current public website (e.g., SVMHC (Donnes and Kohlbacher, 2006)). Additionally, the results imply that combining the analysis of H-2Kb–peptide and peptide–TCR is helpful to identify the potential peptides for H-2Kb.
328
2
I.H. Liu et al.
Methods and materials
Figure 1 summarises the flow chart of methodology in this study. We constructed the peptide query sets from IEDB (Vita et al., 2010) to assess the reliability of template-based scoring function (Figure 1(D)), which we developed successfully in evaluating protein–protein interaction (Chen et al., 2007). Figure 1(C) illustrates that we consider the binding information on both H-2Kb–peptide and peptide–TCR interfaces to investigate the binding mechanism. Figure 1
The flow chart of methodology in this study. (A) The peptide query sets include MHC set and B-MT set. (B) The scoring function is applied on 1G6R, 1KJ2, 1NAM, 2CKB, 2OL3 and 3CVH. (C) Here shows the visualisation of both interfaces. (D) Scoring function and four knowledge-based scoring matrices (see online version for colours)
2.1 H-2Kb–peptide–TCR complex templates This study provides a template-based scoring function to identify the H-2Kb–peptide and peptide–TCR interactions. We collected all co-crystal complexes of H-2Kb, peptide and TCR as structure templates. Here, the H-2Kb–peptide–TCR complex templates, which consist of 6 crystal structures (i.e., 1G6R, 1KJ2, 1NAM, 2CKB, 2OL3 and 3CVH), were extracted from the Protein Data Bank (PDB) (Deshpande et al., 2005) released on 25 July 2010. For each complex template, we identified the contacting residues of two interfaces (H-2Kb and TCR side). Contacting residues, whose any heavy atom should be within a threshold (distance 4.5Å) to any heavy atom of another chain, were considered as the core parts of template-based scoring.
Template-based scoring functions for visualising biological insights
329
2.2 Scoring function and matrices We have recently proposed a template-based scoring function to determine the reliability of a protein–protein interaction and identify the homologous protein complexes (Chen et al., 2007; Lo et al., 2010). In this study, we apply the scoring function to model the H-2Kb–peptide and peptide–TCR binding interface and H-2Kb–peptide–TCR complexes. The scoring function includes the contact-residue interacting score, the template consensus scores and the template similarity scores. The scoring function is defined as Etot = Evdw + ESF + Esim + wEcons
(1)
where Evdw and ESF are the interacting van der Waals energy and the special interacting bond energy (i.e., hydrogen-bond energy and electrostatic energy), respectively; Esim is the template interface similar score; the Econs is couple-conserved residue score. Here, w is set to 3, based on our previous studies of protein–protein interaction. The Evdw and ESF are given as CP
Evdw = ∑ (Vssij + Vsbij + Vsb ji )
(2)
i, j
CP
ESF = ∑ (Tssij + Tsbij + Tsb ji )
(3)
i, j
where CP denotes the number of the aligned-contact residues of peptide p aligned to a hit template; Vssij and Vsbij (Vsbji) are the sidechain-sidechain and sidechain-backbone van der Waals energies between residues i (in H-2Kb) and j (in peptide p), respectively. Tssij and Tsbij (Tsbji) are the sidechain-sidechain and sidechain-backbone special interacting energies between i and j, respectively, if the pair residues i and j form the special bonds (i.e., hydrogen bond, salt bridge, or disulphide bond) in the template structure. The van der Waals energies (Vssij, Vsbi, and Vsbji) and special interacting energies (Tssij, Tsbij and Tsbji) were calculated from the four knowledge-based scoring matrices (Chen et al., 2007; Lo et al., 2010). The value of Esim was calculated from the BLOSUM62 matrix based on the alignments between peptide (p) of the template and their peptide candidates (p'), respectively. The Esim is defined as CP
K jj ′
i, j
K jj
Esim = ∑
(4)
where CP is the number of contact residue pairs in the template; i and j are the contact residue in H-2Kb and peptide, respectively. Kjj' is the score of aligning residue j (in peptide p) to j' (in candidate p') according to BLOSUM62 matrix. Kjj is the diagonal scores of BLOSUM62 matrix for residues j. The couple-conserved residue score (Econs) is determined from two profiles of the template and given by CP
Econs = ∑ (max(0, ( M iS − K ii ) + ( M jS ' − K jj ))) i, j
(5)
330
I.H. Liu et al.
where CP is the number of contact residue pairs; MiS is the score in the PSSM for residue type i at position S in H-2Kb; M jS ′ is the score in the PSSM for residue type j at position S′ in peptide p, and Kii and Kjj are the diagonal scores of BLOSUM62 matrix for residue types i and j, respectively. To evaluate statistical significance (Z-value) of the contact-residue interacting score, the template consensus scores, the template similarity scores, and Etot of a peptide candidate, we randomly generated 5000 interfaces by mutating contact residues of peptide for each H-2Kb–peptide–TCR complexes in template library. The selected residue was substituted for another amino acid residue according to the amino acid composition of Swiss-Prot protein database (Boutet et al., 2007). For all structure templates, we provide an average Z-value, which is the average of all Z-values, as another index for evaluating the significance.
2.3 Data sets To evaluate the scoring function applied to H-2Kb–peptide–TCR complexes, we selected a peptide query set from the IEDB (Vita et al., 2010). IEDB annotates the context in which they are immunogenic and has more entries than any other existing database in this field (Roomp et al., 2010). We collected the H-2Kb/peptide and peptide/TCR from the IEDB. Only the octamer peptides, which were non-modification of residues and interacted with H-2Kb experimentally, were collected. We also filtered the contradiction recorded peptides (i.e., both positive and negative recorded of interacting with H-2Kb or TCR side). Then, the set (termed MHC) has 496 and 525 octamers peptides defined as positive and negative to the H-2Kb, respectively. We selected a sub-set (termed B-MT) from the set MHC. The set B-MT has 86 octamers with both positive recorded on H-2Kb and TCR sides and 67 octamers with both negative recorded on H-2Kb and TCR sides. We compared the performances between the H-2Kb–peptide and peptide–TCR on sub-set B-MT and evaluate the influence of TCR side on MHC side peptide prediction and analysis.
3
Results
On the basis of our previous studies (Chen et al., 2007; Lo et al., 2010), we provide some criteria for using scoring function to ensure the reliability and accuracy of homologous protein–protein interactions. 1
each protein candidate aligned to the template is the homologous of template proteins with the significant sequence similarity (BLASTP E-values ≤10–10)
2
the candidates have ≥25% Sequence Identity (SI) to make sure the binding domains
3
the interfaces have the significant similarity (Z-value ≥3).
The properties of peptides are different from proteins. While we applied the scoring function to protein–peptide interaction such as MHC–peptide and peptide–TCR, we should test the criteria of peptide on the MHC–peptide and peptide–TCR interfaces. We collected all H-2Kb–peptide–TCR complexes from PDB and observe the property of peptides from structure and sequence views. We used the multiple structure
Template-based scoring functions for visualising biological insights
331
alignment tool (CEMC) (Guda et al., 2004) to align the H-2Kb and superimpose the peptides. Figure 2 shows the result of multiple sequence and structure alignment of H-2Kb and peptides. Although the peptides have variant sequence in multiple sequence alignment (Figure 2(B)) and behave diversely by WebLogo program (Crooks et al., 2004) (Figure 2(C)), the backbones of peptides on crystal structure share a highly conserved conformation (Figure 2(A)). The similar conformation of peptides in all crystallographically solved structures of the peptide–MHC complexes is also indicated by previous studies (Madden et al., 1993; Altuvia and Margalit, 2004). On the basis of the conserved conformation, we believe our scoring function could be applied on MHC–peptide and peptide–TCR. Figure 2
The analysis of peptides amount six H-2Kb–peptide–TCR crystal structures by structure and sequence views. (A) Multiple structure alignment of six peptides and H-2Kb. The backbones of six peptides on crystal structure share a highly conserved conformation. (B) Multiple sequence alignment. The peptides have diverse sequence. (C) Sequence logo derived from WebLogo (see online version for colours)
(A)
(B)
(C)
3.1 Evaluation of the scoring function on H-2Kb–peptide interface To evaluate the accuracy of the template-based scoring function on H-2Kb–peptide interface, we predicted peptide candidate from the sub-set B-MT with six templates. Eighty six octamers with both positive recorded on H-2Kb and TCR sides were defined as positive cases and 67 octamers with both negative recorded on H-2Kb and TCR sides were defined as negative cases. Figure 3(A) shows the ROC curves of six templates with original scoring function. We separate the original scoring function (Etot) to observe the major contribution to the MHC–peptide interface (Figure 3(B) and (C)). Although the templates 2CKB and 2OL3 have the worst accuracy of Etot, the template similarities (Esim) still provide a good accuracy (light green and light blue lines in Figure 3(A) and (B)). The predicting accuracy of Esim of six templates is similar and slightly influenced by the SI of candidate peptide (Table 1). However, the predicting accuracy of interacting force (Evdw + ESF) shows extreme divergence from the six templates (Figure 3(C)). On the protein–protein interaction interface, interacting force from our scoring function has an assumption. Candidate proteins, which aligned to the template, should have sequence identities ≥0.25 and they are the homologous proteins of template to keep the similar binding model of the template. On the case of H-2Kb–peptide, the average SI of all peptides in sub-set B-MT is less than 0.25. Template 3CVH with the highest predicting performance has the highest average sequence identity, but otherwise template 2CKB with the worst predicting performance has the worst average sequence identity. Although peptides on the H-2Kb–peptide interface share a conserved backbone conformation, it cannot imply that different amino acids share the similar binding model.
332 Table 1 Template
I.H. Liu et al. The average peptide sequence identities of six H-2Kb–peptide templates Ave. SI (total)
Ave. SI (positive)
Ave. SI (negative)
1g6r
1.24/8
1.55/8
0.85/8
1kj2
1.31/8
1.66/8
0.85/8
1nam
1.15/8
1.48/8
0.73/8
2ckb
1.22/8
1.31/8
1.09/8
2ol3
1.08/8
1.36/8
0.72/8
3cvh
1.4/8
1.69/8
1.03/8
The average sequence identity is computed using 153 octamers (86 and 67 peptides defined as positive and negative cases, respectively) form set B-MT. Figure 3
The ROC curves of different scoring functions based on six H-2Kb–peptide templates. (A) The ROC curves of original scoring function (Etot) on six templates. The templates 3CVH and 2KJ2 have the better AUC than others and 2CKB is the worst template. The separate terms (i.e., template similarity score (Esim). (B) and interacting force (Evdw+ESF). (C)) could provide the detailed influence of template similarity and interacting force. (D) The ROC curves of interacting force (Evdw+ESF) with a Sequence Identity (SI) cut-off ≥1/8 (see online version for colours)
(A)
(B)
(C)
(D)
To ensure the reliability of interacting force (Evdw + ESF) on H-2Kb–peptide interface, we provide a sequence identity threshold (≥1/8). Peptide candidates with sequence identity ≥1/8 are modelled with original interface force. Comparing Figure 3(D) with 3(C), we observed that the ROC curves of interacting force were distinctly reformed
Template-based scoring functions for visualising biological insights
333
by using the cut-off of the sequence identity. Three templates (1G6R, 1NAM and 2OL3) have outstanding improvement and only one template (2CKB) does not. We compare the binding model of the six templates and find out why the template 2ckb has no advantage of interaction force. According to our observation, these templates have the similar number of total interacting residue pairs (49~53 pairs) and 35 of total pairs are conserved on six H-2Kb–peptide interfaces. Figure 4(A) and (B) shows the multiple sequence alignment of the H-2Kb and peptide of six templates. The residues coloured green and grey are forming the hydrogen-bond and van der Waals interaction, respectively. Although the contacting residues of six templates conserved (22 of total 36 positions are consistent with each other), the composition of interacting force have the slight difference. Figure 4(B) indicates the difference between 2ckb and other templates. Template 2ckb lacks of the hydrogen-bond on the end of the peptide. There are two conserved hydrogen-bonds (H-2Kb 77D to peptide Pos.8 and 147W to Pos.7) on other five templates (Figure 4(C)). Figure 4
Multiple sequence and structure alignments of H-2Kb and peptides of six H-2Kb–peptide–TCR complexes. (A) Multiple sequence alignment of six H-2Kb. (B) Multiple sequence alignment of six peptides. The coloured residues mean the interacting residues. All eight residues of the peptides interact with the H-2Kb. The residues coloured green could form the hydrogen-bond interactions. The residues coloured grey only have the van der Waals contacting with other residues. (C) The multiple structure alignment of 1G6R, 1KJ2, 1NAM, 2OL3 and 3CVH. These five peptides have hydrogen-bond on the end (Positions 7 and 8). Residue 77D and 147W of H-2Kb form two conserved hydrogen-bond on positions 7 and 8 of the peptides in five templates. (D) The binding interface of H-2Kb–peptide of template 2CKB. 2CKB lacks of two hydrogen-bonds but only has van der Waals interaction on the end of peptide (see online version for colours)
334
I.H. Liu et al.
On the basis of comparison of multiple templates, we can identify the conserved interacting residue pairs and find which template is different from others or short of the conserved interaction. Discarding the template 2CKB causes a little improvement with predicting accuracy of average Z-value. Finally, we refine our scoring function and apply on the MHC–peptide interaction. The final score is defined as: Final score = Ave.Z vdw + SF + Ave.Z sim
(6)
where Ave.Zvdw+SF is the average Z-value of interacting force (Evdw+ESF); Ave.Zsim is the average Z-value of template similarity (Esim).
3.2 Contribution of peptide–TCR binding for the H-2Kb–peptide–TCR complexes The recognition of foreign antigens by T lymphocytes is essential to most adaptive immune responses. It is driven by specific TCRs binding to antigenic MHC–peptide complex molecules on other cells (Davis et al., 2007). Previous studies (Rammensee et al., 1999; Altuvia and Margalit, 2004; Donnes and Kohlbacher, 2006) only focus on the MHC–peptide interface to predict peptide. In this study, we also dissect the peptide–TCR interface and attempt to provide some important factors of the template-based scoring function. Figure 5(A) and (B) shows the predicting accuracy of two features (i.e., interacting force (Evdw+ESF) and template similarity (Esim)) of scoring function on the sub-set B-MT. Although the binding interface is small (only 9~15 pairs interacting residue pairs), the template similarity still provides a good accuracy. Moreover, template similarity has the similar performance on H-2Kb–peptide side and peptide–TCR side (Figures 3(B) and 5(A)). However, the interacting force has no contribution to predict the interaction peptide–TCR. Figure 5(C) shows that sex templates have the small and diversity interfaces. Additionally, the binding model of each template is also different from others. As stated earlier, interacting force of template-based method has its limitations in modelling the peptide–TCR interaction. Although the scoring function has less predictive accuracy on peptide–TCR, we notice that ROC curve is superlative in MHC–peptide–TCR complexes (Figure 5(D)). This outcome reveals that the interaction of peptide–TCR could promote the recognition of peptide in the MHC–peptide–TCR systems. It implies that our scoring function could provide a distinctive insight into the prediction of recognisable peptides.
Template-based scoring functions for visualising biological insights Figure 5
335
The ROC curves and structures of six templates on peptide–TCR interface and the influence of peptide–TCR interface on MHC–peptide–TCR complex. The separate terms (i.e., template similar score (Esim). (A) and interacting force (Evdw+ESF). (B)) provide detailed influence of template similarity and interacting force on peptide–TCR interface. (C) The contacting residues (coloured residues) of TCR are highly diverse and do not share a similar binding model. Red, blue, green, yellow, purple and light blue are template 1G6R, 1KJ2, 1NAM, 2CKB, 2OL3 and 3CVH, respectively. (D) The ROC curve is more excellent in H-2Kb–peptide–TCR complexes than H-2Kb–peptide or peptide–TCR, respectively (see online version for colours)
(A)
(C)
(B)
(D)
3.3 The performance comparison with other methodologies on the B-MT and MHC sets We compare our score with current public MHC–peptide predicting websites (i.e., SVMHC (Donnes and Kohlbacher, 2006), Predep (Altuvia and Margalit, 2004) and SYFPEITHI (Rammensee et al., 1999)) on the sub-set B-MT and set MHC. Predep also uses the crystal structure as template to model the MHC–peptide interaction. SVMHC is based on the SVM training to predict MHC–peptide interaction. Figure 6 shows the area under the ROC curve (AUC) compared sub-set B-MT with set MHC. Applying to the MHC–peptide interactions, the modification of scoring
I.H. Liu et al.
336
function improves the accuracy (Final bar in Figure 6(A) and (B)). In the sub-set B-MT and set MHC, our scoring function has better performance than Predep, which is a structure-based method. Although SVMHC and SYFPEITHI methods provide better performance than our method on set MHC, our scoring function could suggest the biological insight into MHC-peptide binding mechanism and still possess a reliable predicting accuracy. Figure 6
The AUC of different scoring functions on sub-set B-MT (A) and set MHC. (B) Final means the final score (equation (6); Ave.Zvdw+SF+Ave.Zsim); original scoring function is the original method for protein–protein interactions (equation (1)). SVMHC (Donnes and Kohlbacher, 2006), Predep (Altuvia and Margalit, 2004) and SYFPEITHI (Rammensee et al., 1999) are the public websites for MHC–peptide prediction (see online version for colours)
(A)
4
(B)
Conclusions
This study demonstrates that our template-based scoring functions could enhance the adaptable feasibility for visualising biological insights of H-2Kb–peptide–TCR complexes. The template-similarity (Zsim) has a superiorly predicting accuracy on H-2Kb–peptide and peptide–TCR interactions. Our model provides the detailed interactions and visualisation binding models of MHC–peptide–TCR. Our method can combine with non-template-based servers (e.g., SVMHC) to ensure the predicting accuracy and interaction visualisation. We believe that our model is useful for the development of peptide-based vaccines.
Acknowledgements J.M. Yang was supported by National Science Council and partial support of the ATU plan by MOE. The authors are grateful to both the hardware and software supports of the Structural Bioinformatics Core Facility at National Chiao Tung University.
Template-based scoring functions for visualising biological insights
337
References Aloy, P. and Russell, R.B. (2002) ‘Interrogating protein interaction networks through structural biology’, Proceedings of the National Academy of Sciences of the United States of America, Vol. 99, pp.5896–5901. Altuvia, Y. and Margalit, H. (2004) ‘A structure-based approach for prediction of MHC-binding peptides’, Methods, Vol. 34, pp.454–459. Boutet, E., Lieberherr, D., Tognolli, M., Schneider, M. and Bairoch, A. (2007) ‘UniProtKB/ Swiss-Prot’, Methods Mol. Biol., Vol. 406, pp.89–112. Chen, Y.C., Lo, Y.S., Hsu, W.C. and Yang, J.M. (2007) ‘3D-partner: a web server to infer interacting partners and binding models’, Nucleic Acids Res., Vol. 35, pp.W561–W567. Crooks, G.E., Hon, G., Chandonia, J.M. and Brenner, S.E. (2004) ‘WebLogo: a sequence logo generator’, Genome Res., Vol. 14, pp.1188–1190. Davis, M.M., Krogsgaard, M., Huse, M., Huppa, J., Lillemeier, B.F. and Li, Q.J. (2007) ‘T cells as a self-referential, sensory organ’, Annu. Rev. Immunol., Vol. 25, pp.681–695. Deshpande, N., Addess, K.J., Bluhm, W.F., Merino-Ott, J.C., Townsend-Merino, W., Zhang, Q., Knezevich, C., Xie, L., Chen, L., Feng, Z., Green, R.K., Flippen-Anderson, J.L., Westbrook, J., Berman, H.M. and Bourne, P.E. (2005) ‘The RCSB protein data bank: a redesigned query system and relational database based on the mm CIF schema’, Nucleic Acids Res., Vol. 33, pp.D233–D237. Donnes, P. and Kohlbacher, O. (2006) ‘SVMHC: a server for prediction of MHC-binding peptides’, Nucleic Acids Res., Vol. 34, pp.W194–W197. Guda, C., Lu, S., Scheeff, E.D., Bourne, P.E. and Shindyalov, I.N. (2004) ‘CE-MC: a multiple protein structure alignment server’, Nucleic Acids Res., Vol. 32, pp.W100–W103. Hakenberg, J., Nussbaum, A.K., Schild, H., Rammensee, H.G., Kuttler, C., Holzhutter, H.G., Kloetzel, P.M., Kaufmann, S.H. and Mollenkopf, H.J. (2003) ‘MAPPP: MHC class I antigenic peptide processing prediction’, Appl Bioinformatics, Vol. 2, pp.155–158. Jojic, N., Reyes-Gomez, M., Heckerman, D., Kadie, C. and Schueler-Furman, O. (2006) ‘Learning MHC I peptide binding’, Bioinformatics, Vol. 22, pp.e227–e235. Lin, H.H., Ray, S., Tongchusak, S., Reinherz, E.L. and Brusic, V. (2008) ‘Evaluation of MHC class I peptide binding prediction servers: applications for vaccine research’, BMC Immunology, Vol. 9, pp.8–20. Lo, Y.S., Lin, C.Y. and Yang, J.M. (2010) ‘PCFamily: a web server for searching homologous protein complexes’, Nucleic Acids Res., Vol. 38, Suppl., pp.W516–W522. Lu, L., Arakaki, A.K., Lu, H. and Skolnick, J. (2003) ‘Multimeric threading-based prediction of protein-protein interactions on a genomic scale: application to the saccharomyces cerevisiae proteome’, Genome Research, Vol. 13, pp.1146–1154. Madden, D.R., Garboczi, D.N. and Wiley, D.C. (1993) ‘The antigenic identity of peptide-MHC complexes: a comparison of the conformations of five viral peptides presented by HLA-A2’, Cell, Vol. 75, pp.693–708. Rammensee, H., Bachmann, J., Emmerich, N.P., Bachor, O.A. and Stevanovic, S. (1999) ‘SYFPEITHI: database for MHC ligands and peptide motifs’, Immunogenetics, Vol. 50, pp.213–219. Roomp, K., Antes, I. and Lengauer, T. (2010) ‘Predicting MHC class I epitopes in large datasets’, Bmc Bioinformatics, Vol. 11, p.90. Vita, R., Zarebski, L., Greenbaum, J.A., Emami, H., Hoof, I., Salimi, N., Damle, R., Sette, A. and Peters, B. (2010) ‘The immune epitope database 2.0’, Nucleic Acids Res., Vol. 38, pp.D854–D862.