In silico and cell-based analyses reveal strong divergence between prediction and observation of T-cell–recognized tumor antigen T-cell epitopes

Tumor exomes provide comprehensive information on mutated, overexpressed genes and aberrant splicing, which can be exploited for personalized cancer immunotherapy. Of particular interest are mutated tumor antigen T-cell epitopes, because neoepitope-specific T cells often are tumoricidal. However, identifying tumor-specific T-cell epitopes is a major challenge. A widely used strategy relies on initial prediction of human leukocyte antigen-binding peptides by in silico algorithms, but the predictive power of this approach is unclear. Here, we used the human tumor antigen NY-ESO-1 (ESO) and the human leukocyte antigen variant HLA-A*0201 (A2) as a model and predicted in silico the 41 highest-affinity, A2-binding 8–11-mer peptides and assessed their binding, kinetic complex stability, and immunogenicity in A2-transgenic mice and on peripheral blood mononuclear cells from ESO-vaccinated melanoma patients. We found that 19 of the peptides strongly bound to A2, 10 of which formed stable A2-peptide complexes and induced CD8+ T cells in A2-transgenic mice. However, only 5 of the peptides induced cognate T cells in humans; these peptides exhibited strong binding and complex stability and contained multiple large hydrophobic and aromatic amino acids. These results were not predicted by in silico algorithms and provide new clues to improving T-cell epitope identification. In conclusion, our findings indicate that only a small fraction of in silico–predicted A2-binding ESO peptides are immunogenic in humans, namely those that have high peptide-binding strength and complex stability. This observation highlights the need for improving in silico predictions of peptide immunogenicity.

Tumor exome and transcriptome sequences provide comprehensive information on mutated, overexpressed genes and aberrant splicing, which can be exploited for cancer immuno-therapy. Of special interest are tumor antigen (TA) 3 T-cell epitopes containing mutation(s), because neoepitope-specific T cells often are tumoricidal (1,2). To identify TA-derived T-cell epitopes, MHC-binding peptides are usually identified first. To this end MHC-peptide (pMHC) complexes can be isolated from tumor cells and their peptide cargo sequenced by mass spectrometry (3,4). Alternatively in silico peptide predictions and peptide binding validation are used (5). Modern in silico peptide predictions involve machine-learning techniques like artificial neural networks (ANN) (6). A challenge in peptide prediction is the high diversity of human leukocyte antigen (HLA) alleles; even comprehensive databases, such as IEDB, contain no or limited data for rare alleles, which compromises training of prediction algorithms (7,8). To improve prediction accuracy, pan prediction servers, like NetMHCpan or Pick-Pocket, were introduced that exploit similarities between MHC alleles and their ligand-binding properties (9 -11). Another difficulty is that MHC class I molecules present peptides of different lengths, usually 8 -11 residues long. Gapped sequence alignments were introduced to improve predictions of peptides of different length (9,12,13). By including proteasomal cleavage predictions, peptide prediction accuracy can be further increased (8,14). Different in silico MHC ligand prediction algorithms can be combined to reduce the number of peptide candidates (8,15).
Only a small fraction of HLA ligands is immunogenic, and in silico prediction of these is challenging because of ambiguities of prediction parameters. According to some studies, immunogenicity correlates with peptide binding affinity (5), pMHC complex kinetic stability (16), or both (17). Moreover, it has been reported that immunogenic peptides contain large aliphatic and/or aromatic residues in TCR-accessible positions (18,19). T-cell epitope prediction servers like NetTepi or IEDB immunogenicity integrate such parameters (18,20). However, peptide immunogenicity also depends on other factors, such as the efficiency of their production and presentation by professional antigen-presenting cells (APCs) and on central tolerance (1,2). For personalized cancer immunotherapy, it is crucial to identify TA-specific T-cell epitopes, and the available procedures are error-prone (21).
To identify key parameters of CD8 ϩ T-cell epitopes, we used the cancer testis antigen NY-ESO-1 (ESO) and HLA-A*0201 (A2). This non-mutated TA is expressed on a wide range of tumors, is highly immunogenic, and has been used in diverse vaccine studies, and CD8 ϩ T-cell responses have been studied extensively (22)(23)(24)(25)(26)(27). Four A2-restricted ESO epitopes have been described that are naturally produced and presented by APC, two of which are expressed by tumor cells (25, 28 -31). ESO-specific CD8 ϩ T-cell responses in humans exhibit a strong immunodominance hierarchy and diverse HLA restrictions (32,33). Here we used different in silico servers to predict the binding strength, complex kinetic stability, and immunogenicity of A2-restricted 8 -11-mer ESO peptides. The 41 peptides with the highest predicted binding affinity were tested for (i) binding to A2 using a refolding, a peptide rebinding assay (34,35), and an A2 stabilization assay on T2 cells (36); (ii) A2-peptide complex kinetic stability at 37°C (34,35,37); (iii) peptide immunogenicity in A2/DR1 transgenic H-2 Ϫ/Ϫ mice (36, 38 -40); and (iv) recognition by CD8 ϩ T cells from ESOvaccinated melanoma patients (29 -31). Our results define parameters of peptide immunogenicity and provide new cues on how to improve T-cell epitope discovery.

ESO peptide binding to A2
To predict the binding of ESO 8 -11-mer peptides to A2, we used the NetMHC 3.4 server (5,(41)(42)(43). By setting an affinity threshold of 3000 nM, 41 peptides were obtained (Table 1). Of these, 19 were 10-or 11-mers, which was unusual, because normally the majority of A2 bound peptides are 9-mers (13). The immunodominant ESO 157-165 peptide had an IC 50 of 1015 nM and would have been missed when using the recommended cutoff of 500 nM (28,29). We also performed predictions using the IEDB MHC I prediction server (44) and obtained the same results plus 17 additional peptides with predicted IC 50 values of 918 -2700 nM, none of which have been reported previously (supplemental Table S1).
Binding of the peptides to A2 was measured in a refolding assay (45). The most efficient refolding was observed for peptide 4, referred to as 100%, followed by peptides 6 (90%), 5 (87%), 31 (83%), and 1 (80%) (Fig. 1, A and B). The correlation between measured and predicted peptide binding exhibited a Pearson coefficient of r ϭ 0.64 and strong divergences for the peptides with high refolding scores (supplemental Fig. S1A). Similar correlations were observed when peptide binding was predicted with the more recent NetMHC 4.0 or NetMHCpan servers (9, 12) (supplemental Fig. S1, B and C). Comparable binding values were observed when using a peptide-rebinding assay (r ϭ 0.97) (Fig. 1, B and C). Repeating these experiments with different batches of peptides cautions that errors in the peptides (e.g. impurities and degradations) can be larger than those of these assays. We also assessed ESO peptide binding by an A2 complex stabilization on A2 ϩ , TAP Ϫ T2 cells (36) and obtained grossly different results, which may be explained by the fact that A2 peptide stabilization on these cells relies on different mechanisms, the relative contributions of which are peptide-dependent (data not shown) (46).

A2-ESO peptide complexes kinetic stability
The kinetic stability of the A2-ESO peptide complexes obtained in Ͼ30% yields was assessed at 37°C, and their halflives (1 ⁄ 2 ) were calculated ( Fig. 2A). Of the 20 complexes analyzed, the 1 ⁄ 2 ranged between 1.82 h (peptide 29) and 13.5 h (peptide 1). The most stable complexes were those containing the peptides 1, 4, 16, and 31 and the Flu matrix 58 -66 peptide. These peptides also exhibited strong A2 binding (Fig. 1A); however, other peptides exhibited robust A2 binding, but low kinetic complex stability, and the overall correlation between measured A2 binding and complex kinetic stability was poor (r ϭ 0.54) (Fig. 2B). The measured complex stabilities correlated even less well with those predicted by the NetMHCstabpan (9) or the NetMHCstab server (47) (r ϭ 0.31 and 0.47) (supplemental Fig. S1, D and E). It noteworthy that the correlation between predicted complex stabilities and predicted binding affinities was better when using the NetMHC 3.4 rather than the more recent NetMHC 4.0 server (r ϭ 0.58 and 0.33, respectively) ( Fig. 2C and supplemental Fig. S1F).

Identification of cancer T-cell epitopes ESO peptide immunogenicity
To assess the ESO peptide immunogenicity in mice, groups of A2/DR1 transgenic H-2 Ϫ/Ϫ animals were immunized with pools of five peptides of comparable A2 binding affinity. Fourteen days after a booster immunization, CD8 ϩ T-cell splenocytes were isolated and tested for IFN␥ production by ELISPOT upon incubation with single peptide pulsed T2 cells. For the 10 peptides 1-4, 6, 8, 14, 16, 31, and 32, IFN␥ responses were observed in the range of 30 -108 spots/10 5 T cells (Fig. 3A). The strongest responses were observed for peptides 1, 4, 8, 16, and 32. To assess peptide immunogenicity in humans, purified CD8 ϩ T cells from two melanoma patients vaccinated with recombinant vaccinia and fowl pox vectors expressing fulllength ESO (23) were stimulated with the ESO peptides and assayed for IFN␥ ELISPOT upon incubation with ESO peptide pulsed T2 cells. Strong IFN␥ responses (600 -800 spots/10 5 T cells) were observed on the cells from patient NW 1789 for peptides 1, 4, 6, 16, and 31 ( Fig. 3B and supplemental Fig. S2A, blue bars). Lower responses were observed when autologous DC were used as APC ( Fig. 3B and supplemental Fig. S2A, red bars). The peptide-dependent variations of the reductions may be explained by biased peptide presentation by T2 cells; e.g. the peptides 6 and 31 had higher binding scores on T cells than the peptides 1, 4, and 16. It may also be that on DC some peptides are presented by HLA alleles other than A2, which on T2 cells is unlikely, because they express A2 and only scant levels of HLA-B51 and Cw1 (supplemental Fig. S2B) (48). For the peptides 4, 16, and 31 CD8 ϩ T-cell responses have been described previously (supplemental Fig. S3). The peptide ESO 155-163 was missed, because its predicted binding affinity was 3319 nM, i.e. above the cutoff of 3000 nM used. The A2-restriced CD8 ϩ T-cell responses for peptides 1 and 6 have not been reported previously. Remarkably, ESO peptides 2, 3, 8, and 14 were immunogenic in A2 transgenic mice, but not in humans (Fig. 3).

Parameters defining ESO peptide immunogenicity
The peptides that were immunogenic in humans exhibited the highest A2 binding and kinetic complex stability (Fig. 2B). For the peptides immunogenic in mice only, both parameters were slightly lower. All immunogenic peptides exhibited complex stabilities of Ͼ4 h and refolding scores of Ͼ50%, and all non-immunogenic peptides exhibited lower values. No such correlation was observed when peptide-binding affinity was predicted using the NetMHC 3.4 (6, 42), NetMHC 4.0 (12), or NetMHCpan (9) server or kinetic A2-ESO peptide complex stability using the NetMHCstab (47) or NetMHCstabpan (16) server ( Fig. 2C and supplemental Fig. S1). However, for most of the ESO peptides, the T-cell epitope scores predicted by the NetTepi server (20) correlated better with the measured bind- Only the sequence of the ESO 86 -94 peptide was outside these two registers (Figs. 3 and 4, C and E). When bound to A2, generally the side chains of the second and the last (C-terminal) residues occupy the B and F pockets, whereas the side chains of the others are solvent-exposed to different degrees, and some can be secondary anchor residues (49 -51). The peptides comprising the ESO 159 -165 core sequence exhibited four or five large aliphatic and/or aromatic residues in these positions, whereas the peptides containing the ESO 110 -116 core sequence only one or none (Fig. 4D). Two studies have shown that immunogenic peptides express such amino acids in solvent-exposed positions (18,19). Indeed, the immunogenicity scores calcu-lated with the IEDB immunogenicity predictor (http://tools. iedb.org/immunogenicity/), 4 which takes TCR propensity into account, were higher for the peptides containing the ESO 159 -165 core sequence (0.17-0.25) than those containing the ESO 110 -116 sequence (Ϫ0.06 -0.009) ( Table 1).
The correlations between predicted and measured peptide binding and kinetic pMHC complex stability in our study were poorer compared with those reported in other studies (supplemental Fig. S1, A-E) (9,16,42,43). In these studies and for the training of the prediction servers, pathogen-derived antigens were used. To address the question of whether there are differences between TA and pathogen-derived peptides, we examined 149 non-mutated TA and 129 viral T-cell epitopes. All peptides were A2-restricted nonamers and collected from databases (supplemental Table S2). Positional amino acid usage of these peptides was analyzed with the Seq2Logo server (52). This revealed that in the main A2 anchor positions, 2 and 9, Leu was

Identification of cancer T-cell epitopes
more frequent in TA peptides, especially in P9, in which Val was the most abundant residue in viral peptides (supplemental Fig.  S4, A and B). Moreover, viral peptides exhibited higher amino acid diversity especially in positions 3 and 7, which typically are secondary A2 anchor residues (50,51). We next calculated the average hydrophobicity scores for the residues in positions 1-9 for the two sets of peptides using the scales published by Kyte and Doolittle (53). The hydrophobicity was highest for the residues in position 2 and 9 and lowest for those in position 4 (supplemental Fig. S4C). Similar results were obtained when other amino acid hydrophobicity scales were used, i.e. those determined by Hopp and Woods (70), Abraham and Leo (71), Black and Mould (72), Sweet and Eisenberg (73), and Roseman (74), as detailed in http://web.expasy.org/protscale/ (supplemental Fig. S4D). The TA peptides exhibited higher hydrophobicity in all positions, except for position 3. In position 4 viral but not TA peptide frequently contained acidic residues, resulting in greatly reduced average hydrophobicity. As illustrated in structure of the A2-HIV RT 309 -317 complex, an acidic residue in position 4 can stably bind to the A2 ␣1 helix (Arg-65) (51). Moreover, viral peptides contained more polar and/or charged residues in positions 1 and 7 than TA peptides, accounting for their overall modestly reduced hydrophobicity. Collectively these results argue that at large there exist differences between TA and viral peptides, notably in amino acid usages in A2 anchor positions.

Discussion
A widely used strategy to identify T-cell epitopes consists of first predicting HLA binding peptides by in silico algorithms (13,34,35,54). Here we predicted A2-restricted 8 -11-mer peptides of ESO using the NetMHC 3.4 and IEDB servers and obtained partially overlapping results, which was explained by that these servers are based on related ANN (Table 1 and supplemental Table S1) (6,41,42,44). Testing of the 41 peptides with the highest predicted binding affinity gave very similar results when using the refolding or peptide rebinding assay (Fig. 1).
There exist diverse in silico MHC-peptide binding predictors of which the ANN based NetMHC servers performed best in benchmark studies (8,43,55). We examined correlations between measured A2 ESO peptide binding and NetMHC 3.4 and the more recent NetMHC 4.0 or NetMHCpan 3.0 servers that allow insertions and deletions in peptide alignments and integration of multiple receptor and peptide length data sets, respectively (6,9,12,42). Surprisingly, these refinements did not improve the correlations between measured and predicted ESO-peptide binding (supplemental Fig. S1, A-C). Poorer correlations were observed when peptide binding was predicted with the PickPocket, SYFPEITHI, or Rankpeptide servers, which is consistent with other reports (8,11,43,55). 5 One explanation for the modest correlation between measured and predicted ESO peptide binding could be that the predicted binding affinities represent IC 50 values in nM, whereas in our study relative binding values were measured at a fixed peptide concentration. However, in our system the measured IC 50 val-ues correlated even less well with NetMHC 3.4 predicted values (r ϭ 0.56, p Ͻ 0.0001). 5 We next measured the A2-ESO complex stabilities and found that the measured values poorly correlated with those predicted by the NetMHCstab and the more recent NetMHCstabpan server (supplemental Fig. S1, D and E) (16,47). The measured ESO peptide binding and kinetic complex stability exhibited a modest correlation (Fig. 2B), which is consistent with the fact that peptide binding and kinetic pMHC complex stability can diverge; it has been shown that for complex stability, but much less for peptide binding affinity, fitting of the side chains of the residues in position 2 and 3 into HLA binding pockets is critical (37,47,56). In this plot all immunogenic peptides exhibited complex stabilities of Ͼ4 h and peptide binding efficiencies of Ͼ50% (Figs. 2B and 3 and Table 1). The highest peptide-binding strengths and complex stabilities were observed for the ESO peptides immunogenic in humans, followed by those immunogenic in A2 transgenic mice only. Our results are in accordance with an analysis of large data sets at left indicate the peptide number, those at right indicate the peptide length, and the red stars mark previously reported immunogenic peptides. D, the immunogenic peptides are represented with the potentially solvent-exposed amino acids in bold; highlighted in yellow are large hydrophobic residues, in magenta are the aromatic residues, and in gray are the main A2 anchor residues. E, the ESO sequence with the three immunogenic core sequences highlighted as in C. The residues shown in underlined red indicate proteasomal cleavage sites as predicted by the NetChop 3.1 server.

Identification of cancer T-cell epitopes
showing that immunogenic peptides typically exhibit high binding affinity and high complex kinetic stability (57) but are at variance with studies reporting that peptide immunogenicity correlates with binding affinity (5,34,58) and pMHC complex kinetic stability (16,34,37), respectively. No significant correlation was observed between peptide immunogenicity, binding strength, and complex kinetic stability when complex stability was predicted by the NetMHCstabpan and binding affinity by the NetMHC 3.4 or NetMHC 4.0 server (Figs. 2C and 3 and supplemental Fig. S1F). There was also no correlation between peptide immunogenicity and predicted binding affinity and kinetic complex stability, respectively ( Fig. 3 and supplemental  Fig. S1, A-E, and Table 1). Moreover, our results caution that selection of peptides based in silico predictions using a cutoff affinity of 500 nM is prone to miss immunogenic peptides: in our study three peptides, including the clinically important peptide 31 (supplemental Fig. S3 and Table 1). The same is true for other clinically important TA epitopes like the A2-restricted Melan-A 26 -35 (EAAGIGILTV) (5164 nM), survivin 96 -104 (LTLGEFLKL) (2002 nM), or CEA 694 -702 (GVLVGVALI) (833 nM) peptides (http://www.iedb.org). 4 It is important to note that CTL tumor control depends more on the affinity of pMHC-TCR than on MHC-peptide binding (59).
The correlations between measured and predicted peptidebinding strength and kinetic complex stability were poorer in our study compared with those in other studies (supplemental Fig. S1, A-E) (9,16,42,43). In these and for training of the prediction servers, pathogen-derived antigens were mainly used. Tumor antigens excluding neoantigens are a priori selfantigens and hence are subject to central tolerance, which is not the case for pathogen-derived antigens (60,61). By comparing 149 TA and 129 viral A2 restricted nona-peptides, we observed significant differences in amino acid usages and average hydrophobicity in the potential secondary anchor residues in positions 1, 4, and 7 and smaller ones in the potential main A2 anchor residues in positions 2 and 9 (supplemental Table S2 and Fig. S4). It has been demonstrated that changes in HLApeptide anchoring can alter the conformation and flexibility of pMHC complexes and thus their interaction with TCR (62)(63)(64)(65). The significance of such changes is illustrated, for example, by modification of a potential main anchor residue (A27L) in the Melan-A 26 -25 peptide, which resulted in different TCR interactions and different outcomes of vaccine trials (64,65). Thus the limited in silico prediction accuracy of TA peptides may be explained by that the servers were trained on pathogen-derived peptides. It is noteworthy that a substantial fraction of neoepitopes contains a mutation in an MHC anchor position, some of which may affect T-cell recognition via structural changes in the pMHC complex and binding to TCR (1,2,15,58,62,63). It should be mentioned that most neoepitopes contain a stochastic somatic mutation, which can have different and diverse effects, making general predictions difficult.
An unexpected finding was that immunization of A2 transgenic mice induced CTLs for 10 of the ESO peptides, whereas in humans T cells specific for only five were observed (Figs. 3 and supplemental Figs. S2 and S3). This observation cautions that peptides can be immunogenic in HLA transgenic mice but not in humans. All ESO peptides that were immunogenic in humans contained the sequence ESO 159 -165 (LMWITQC) (Figs. 3 and 4, C-E). When bound to A2, these peptides contained amino acids with potentially solvent-exposed large hydrophobic and/or aromatic side chains, which have been shown to convey immunogenicity (18,19). Conversely, peptides 2, 8, 14, and 32 that were immunogenic only in A2 transgenic mice contained the ESO 110 -116 sequence (AQDAPPL). When bound to A2, these peptides contained one or no such residue (Fig. 4, C-E). In accordance with this, the immunogenicity scores predicted by the IEDB immunogenicity server, which considers the TCR propensity of the peptide, were substantially lower for these than the former peptides (Table 1) (18).
However, peptide 3 contained several hydrophobic/aromatic residues, had high immunogenicity scores, yet was not immunogenic in humans and only weakly in A2 transgenic mice, arguing that immunogenicity also depends on other factors, such as: (i) in humans, but not in HLA transgenic H-2 Ϫ/Ϫ mice, ESO peptides can be presented and recognized in the context of other HLA alleles; e.g. HLA-B35 and Cw3 for which immunodominant ESO CTL responses are known (27,32,33); (ii) the efficiency of peptide production and presentation by APC: in silico predictions and in vitro digestion experiments argue that human proteasomes produce the peptides that were immunogenic in mice, but not in humans (Fig. 4E) (14,32,66); indeed, CTLs were found in cancer patients with such specificities but other HLA restrictions (32,33,67); (iii) peptides binding to multiple HLA alleles, including HLA class II molecules are more immunogenic than those binding to only one allele (32,40,68); for the ESO 159 -165 core sequence-containing peptides, there is the strongly immunogenic, DP4-restricted T-cell epitope ESO 157-170 (69), whereas for the ESO 110 -116 core sequence containing peptides no CD4 ϩ T-cell epitope is known; and (iv) non-mutated TAs, including ESO, are self-antigens, and therefore TA-specific T cell responses are pruned by central tolerance in humans, which is not the case in mice that lack ESO (22,60,61).
In conclusion, our study demonstrated that only a small fraction of A2 binding ESO peptides was immunogenic in humans, namely those that had high peptide-binding strength and kinetic complex stability. These peptides contained multiple hydrophobic/aromatic residues, supporting the notion that immunogenicity correlates with TCR propensity. There is a need to improve in silico predictions of peptide-binding properties and immunogenicity of TA, namely by considering structural/conformational aspects of MHC-peptide binding, training of prediction servers with TA peptides, and refining TCR propensity calculations.

Peptides
Peptides were produced by the Protein and Peptide Chemistry Facility of the University of Lausanne, HPLC-purified (Ͼ95% pure), verified by mass spectrometry, and kept lyophilized at Ϫ80°C.

Peptide-driven refolding assay
Refolding with A2 heavy chain carrying a C-terminal BirA substrate peptide, Cy5-labeled ␤2m, and a test peptide were performed essentially as described (45). Human ␤2m was mutated Ser-88 to Cys and after refolding alkylated with maleimide-PEG 2 -Cy5 (Pierce, Thermo Fisher Scientific) in PBS at pH 7.4. Refolding reactions were performed in 96-well plates at 4°C for 72 h in the presence of 10 M peptide. Incubation without peptide and with the Flu matrix 58 -66 peptide served as negative and positive controls, respectively. After centrifugation (4,000 rpm, 5 min), the reaction mixtures were transferred into 96-well plates, and Cy5 fluorescence was read on a fluorescence plate reader (Modulus, Promega). All measurements were performed in triplicate and data processed using Excel (Microsoft).

Peptide rebinding and pMHC complex kinetic stability assays
96-well plates were coated with streptavidin, and biotinylated A2-MelanA 26 -35 complexes (1 g/ml) were added in 50 l and incubated for 2 h at 4°C. The plates were saturated with biotin, washed, and incubated for 3 min at 4°C with citrate buffer (0.13 M citric acid, 66 mM Na 2 HPO 4 , 150 mM NaCl, pH 4.0), followed by washing with PBS containing 0.05% Tween 20. Test peptides (10 M) and Cy5-labeled ␤2m were added in PBS containing 5 mM EDTA, and the plates were incubated at 4°C for 72 h. After washing with PBS containing 0.05% Tween 20, the A2-peptide complexes were quantified as described above. Incubations without peptide and with the Flu matrix 58 -66 peptide served as negative and positive controls, respectively. The kinetic stability of complexes was assessed by incubating the A2-peptide complexes at 37°C, and after different period of times their content was quantified likewise. The results were plotted, and half-lives were determined using GraphPad Prism software (GraphPad, San Diego, CA). All measurements were performed in triplicate.

Immunization of A2 transgenic mice
HLA-A2/DR1 transgenic, H-2 Ϫ/Ϫ mice (38) were obtained from Taconic Models for Life, maintained in the institute's animal facility, and used in accordance with the Cantonal Veterinary Office. Groups of mice (n ϭ 5) were immunized with peptides essentially as described (39). In brief, pools of five ESO peptides of similar affinities for A2 and the DR1 restricted influenza HA 306 -318 peptide (10 g each) were injected subcutaneously at the base of the tail in an emulsion containing PBS, incomplete Freund's adjuvant (IFA), and oligodinucleotides (ODN) 1826 (InvivoGen, San Diego, CA). After 2 weeks, the mice were booster-immunized, and a fortnight later their spleens were harvested, and the CD8 ϩ T cells were purified by negative selection (Stemcell Technologies, Köln, Germany) and incubated overnight with T2 cells previously pulsed with 1 M of peptide at a 1:1 ratio. Production of IFN␥ was assessed using a mouse ELISPOT kit following the manufacturer's instructions (Mabtech, Nacka Strand, Sweden).

Analysis of NY-ESO-1-specific CTLs from patient
Isolation, culturing, and stimulation of PBMC and CD8 ϩ T cells from melanoma patient NW1789 and NW3276 followed established procedures (31). In brief, purified CD8 ϩ T cells were stimulated twice with 1 M of ESO peptides, irradiated autologous PBMC, and 150 units/ml of IL2. After a fortnight, the CTLs were tested for IFN␥ production by ELISPOT following incubation with ESO peptide-pulsed T2 cells or autologous DC. A positive response was considered if the number of spots in the peptide-exposed well was Ͼ2-fold higher than the number of spots in the unstimulated well, and there were Ͼ10 specific spots/25,000 T cells. The generation of DC and the ELISPOT assay were performed as described (31).

Statistics
Statistical analyses were performed using the GraphPad Prism software (GraphPad). Correlation analyses were performed using Pearson coefficient r. The associated p value (two-tailed, ␣ ϭ 0.05) quantifies the likelihood that the correlation is due to random sampling.
Author contributions-J. S. and P. G. performed the biochemical assays, which were established and optimized by D. D.; J. K. performed all the experiments on human cells; J. S. performed in silico predictions, data processing, and statistical analysis; I. L. and G. C. coordinated the study and edited the manuscript; and all authors discussed and interpreted the results.