Identification of Noncanonical Melanoma-Associated T Cell Epitopes for Cancer Immunotherapy1
http://www.100md.com
免疫学杂志 2005年第11期
Abstract
The identification of tumor-associated T cell epitopes has contributed significantly to the understanding of the interrelationship of tumor and immune system and is instrumental in the development of therapeutic vaccines for the treatment of cancer. Most of the known epitopes have been identified with prediction algorithms that compute the potential capacity of a peptide to bind to HLA class I molecules. However, naturally expressed T cell epitopes need not necessarily be strong HLA binders. To overcome this limitation of the available prediction algorithms we established a strategy for the identification of T cell epitopes that include suboptimal HLA binders. To this end, an artificial neural network was developed that predicts HLA-binding peptides in protein sequences by taking the entire sequence context into consideration rather than computing the sum of the contribution of the individual amino acids. Using this algorithm, we predicted seven HLA A*0201-restricted potential T cell epitopes from known melanoma-associated Ags that do not conform to the canonical anchor motif for this HLA molecule. All seven epitopes were validated as T cell epitopes and three as naturally processed by melanoma tumor cells. T cells for four of the new epitopes were found at elevated frequencies in the peripheral blood of melanoma patients. Modification of the peptides to the canonical sequence motifs led to improved HLA binding and to improved capacity to stimulate T cells.
Introduction
The identification of tumor-associated Ags (TAA)5 and tumor-associated T cell epitopes (TATE) has led to new therapeutic vaccination strategies for the treatment of cancer (1, 2, 3) and is the basis for monitoring tumor-specific immune responses in patients (4). Since the mid-1990s, cancer vaccines of various designs such as synthetic peptides administered alone or together with IFAs or IL-2, or loaded onto autologous dendritic cells were tested in clinical trials (5, 6, 7, 8, 9). Although some trials have shown promising clinical responses, these were usually seen in no more than 20% of the patients. This low efficacy of therapeutic vaccination against cancer may, among other possible reasons, be caused by anergy, low functional avidity, or low frequency of the tumor-specific T cells (10, 11), or by a selection of Ag-loss variants of the tumor cells (12). It is now generally accepted that anti-tumor vaccines should include many different T cell epitopes to reduce the risk of immune evasion, and to address and activate as many tumor-specific T cells as possible. Consequently, the search for new TATE is greatly intensified. The most widely applied strategy for the identification of T cell epitopes, dubbed "reverse immunology," uses bioinformatic tools to predict peptides from the sequences of TAA that have the capacity to bind to the HLA molecules of interest. The predicted epitopes then need to be synthesized, tested in T cell assays, and validated as naturally processed and presented by the tumor cells (13, 14). The bioinformatics is based on HLA allele-specific sequence motifs that correspond to the amino acids that anchor the peptides in the peptide-binding grooves of the HLA molecules (15, 16, 17). The predictions use algorithms that are trained with the frequencies of the different amino acids at the different sequence positions of an epitope (18) or with binding data of known epitopes (19). As an example, HLA-A*0201, the most frequent HLA class I allomorph in Caucasian populations, binds preferentially nonapeptides with the aliphatic amino acids leucine, isoleucine, valine, or methionine at positions 2 and 9. The majority of the T cell epitopes known to date have been identified with such bioinformatic approaches. Notwithstanding, the available algorithms are very restrictive in that they select epitopes for their HLA-binding capacity and base these predictions predominantly on the individual input of the anchor amino acids. However, T cell epitopes need not be good HLA-binding peptides, as continuous production by the Ag-processing machinery may ensure sufficient representation at the cell surface for T cell stimulation. Moreover, it is expected and has been shown that, in HLA-restricted T cell epitopes, the entire sequence context and not only the amino acid occupancy at the individual sequence positions determines the HLA-binding and T cell-stimulating capacity of the peptides (20, 21). To overcome the limitations of the available bioinformatic tools, we developed a prediction algorithm that uses multilayer artificial neural networks (ANN) that, by their parallel data processing, compute the HLA-binding properties of peptides based on the entire sequence context and, thus, can account for the interdependences of the individual amino acids in an epitope. To test this algorithm we predicted HLA A*0201-restricted epitopes of the melanoma-associated TAAs gp100, MAGE-A1, MAGE-A2, p53, and tyrosinase-related protein 2 (TRP-2), and chose for further analyses seven of the predicted peptides that do not conform to the HLA allele-specific epitope sequence motifs for this HLA allomorph.
Materials and Methods
Intracellular cytokine (ICC) staining
ICC staining for epitope-specific CD8+ T cells in the peripheral blood of melanoma patients was described elsewhere (27). Briefly, PBMC were used after cryopreservation and incubated for 30 min at RT and 30 min at 37°C in DMEM (Invitrogen Life Technologies) with 0.1% BSA (Sigma-Aldrich) with the peptides at a concentration of 10 μg/ml. Then, equal volumes of 20 μg of BFA per milliliter of DMEM with 20% FCS was added, and the incubation continued for 5 h. The cells were pelletted, resuspended in PBS/1 mM EDTA, and incubated for 10 min at 37°C. After centrifugation and resuspension, lysis buffer (BD Biosciences) was added, the cells incubated for 3 min at RT, then washed and incubated with permeabilization buffer (BD Biosciences) for 10 min at RT in the dark. For flow cytometry, the cells were stained with anti-CD8-allophycocyanin, anti-CD3-PerCP, anti-CD69-PE, and anti-IFN--FITC (all BD Biosciences) and analyzed with a FACSCalibur flow-cytometer and CellQuest software (BD Biosciences).
Results
Prediction of noncanonical T cell epitopes
This study aimed at identifying with bioinformatic means HLA class I-restricted TATEs that bind only weakly to the corresponding HLA molecules and, therefore, are missed by other algorithms that are designed to predict peptides that bind strongly to HLA molecules and that conform to the HLA allele-specific epitope motifs as defined by the canonical amino acids listed in the SYFPEITHI database (Ref. 12 ; see introduction). To identify epitopes independent from these restrictive criteria we developed a strategy of in silico pattern recognition of HLA-binding T cell epitopes by ANN, which by their parallel mode of data processing can extract common features of complex amino acid sequence patterns (22, 23). Besides the specific design of the ANNs and the selection of input data sets, the predictions depend critically on the descriptors used to code the amino acids. We used the physicochemical-properties hydrophobicity, polarity, refractivity, and side chain volume as referenced (28, 29, 30). The ANNs were applied to MAGE-A1 and A2, gp100, p53, and TRP-2 to identify HLA A*0201-restricted epitopes. The output data included the known T cell epitopes gp100178 MLGTHTMEV, gp100619 RLMKQDFSV; TRP-2180 SVYDFFVWL, TRP-2288 SLDDYNHLV, TRP-2360 TLDSQVMSL; and MAGE-A2112 KMVELVHFL that feature canonical anchors as well as gp100154 KTWGQYWQV, gp100280 YLEPGPVTA, gp100209 ITDQVPFSV, gp100639 RLPRIFCSC; and MAGE-A1278 KVLEYVIKV that deviate from these motifs (31, 32, 33). The seven newly predicted epitopes together with variants with the canonical leucine at the anchor positions 2 and 9 were studied for their capacity to bind to HLA-A*0201 and to stimulate specific T cells responses.
Table I lists the 7 predicted epitopes and the variants with leucines at positions 2 and 9, together with the scores calculated by our ANN. The algorithms had been trained with peptide sequences classified as either HLA-A*0201 ligands or as negative for HLA A*0201-binding. Correspondingly, the output defines the test sequences as potential epitopes or as negative without grading, meaning that any number in Table I column ANN is a positive output. Thus, all 7 peptides and their variants were classified as potential epitopes despite differences in the scoring. The 14 peptides were also scored with the algorithms from the BioInformatics and Molecular Analysis Section (BIMAS) of the Center for Information Technology, National Institutes of Health, which is based on kinetic parameter of peptide-binding to the HLA molecules and predicts the half-life times of the respective HLA peptide complexes (19). Here, only the peptide TRP-2185 is identified as a potential epitope, none of the other 6 peptides are (Table I). In contrast, all of the variants peptides with the motif amino acid leucine in the anchor positions were scored highly. Consequently, TRP-2185 had been identified earlier with BIMAS (34), but none of the other peptides. Also the SYFPEITHI algorithm, which scores peptides according to the degree to which they conform to the HLA allele-specific epitope motifs indicates improved HLA-binding properties of the modified peptides (18). The SYFPEITHI scoring system values the amino acids in specific sequence positions that are conserved among different epitopes, and sums up these individual values to assess the capacity of the peptide to bind to the HLA molecule. Canonical amino acids in the dominant anchor positions get a value of 10. Because there are two dominant anchor positions that define the HLA allele-specific epitope motifs, strong binders should have a score of at least 20. A value below 20 indicates that at least one anchor position carries a suboptimal amino acid. Therefore, 20 is usually taken as the cut-off for T cell epitope prediction. By this criterion, none of the 7 peptides selected by our approach would qualify. However, the modified peptides do so as they were purposely adapted to the canonic anchor motifs (Table I). For an experimental assessment of the capacity of the 14 peptides to bind to HLA-A*0201, we performed HLA stabilization assays with the TAP-deficient cell line T2. Because of the TAP defect these cells lack the proper peptides for binding to and stabilizing of the native conformation of the HLA class I molecules. As a consequence, the cell surface expression levels of HLA are low but can increase when the cells are incubated with suitable peptides. The levels of HLA expression detected after peptide-pulsing can then be used as measure for the strength of binding as illustrated for the influenza matrix protein peptide MP58 in comparison with T2 cells without peptide (Fig. 1). Except for TRP-2185, HLA-A*0201 expression levels after incubation with the original peptides were unaltered when compared with the background without peptide. For TRP-2185 that already in its unmodified form scored highly in BIMAS and SYFPEITHI predictions, the HLA expression after incubation with the natural epitope was high above background but still topped by the introduction of leucine into the anchor positions. For MAGE-A1237 and gp100286, the modifications resulted in peptides that, in contrast to their natural counterparts, stabilized the HLA molecules indicating enhanced binding. Consistent with the predictions by the BIMAS algorithms, the modification of peptide p53347 did not produce a peptide with good HLA-binding properties. These results prove that six of the seven peptides predicted by our algorithms are weak HLA binders and would not be predicted as T cell epitopes by the other algorithms.
After in vitro priming and restimulation of CD8+ T cells from the peripheral blood of five healthy HLA A*0201-positive donors we could detect by ELISPOT assays specific responses against all seven epitopes. MAGE-A2116, MAGE-A1237, p5317, p53347, and TRP-2185 induced strong response (Table III). For MAGE-A2250 and gp100286 strong responses were seen only against the modified but not the natural epitopes. Two of the donors responded to six epitopes each, one against three and two against five epitopes. Taking the two sets of analyses with PBMC from melanoma patients and healthy donors together, all of the seven predicted noncanonical HLA-binding epitopes did induce specific T cell responses, six of them (MAGE-A2116, MAGE-A1237, gp100286, p5317, p53347, and TRP-2185) strong responses. Five of these (MAGE-A2116, MAGE-A1237, MAGE-A2250, gp100286, and TRP-2185) had already primed high frequencies of T cells in the patients. TRP-2185 had been identified before as an HLA A*0201-binding peptide but not yet as a T cell epitope (34). MAGE-A2116 was reported earlier as an HLA A3-restricted epitope (36). We show in this study for the first time that TRP-2185 is indeed a T cell epitope and that MAGE-A2116 is, in addition to HLA-A3, presented by HLA-A*0201. Six of the seven new HLA A*0201-restricted T cell epitopes were not identified by any of the other prediction algorithms for strong HLA-binding epitopes.
Comparison of the natural and the modified versions of the predicted epitopes
Several epitopes were recognized by the T cells in their natural as well as modified form. Therefore, we tested by ELISPOT whether T cells primed with the modified versions of the epitopes TRP-2185, MAGE-A2116, or MAGE-A2250 also respond to the natural variant of these peptides. In all cases, the CD8+ T cells primed with the modified epitope responded to both versions (Fig. 3A). In the case of TRP-2185, the responses to the natural epitope were higher, whereas for the other two, the T cells responded better to the priming peptide than to their natural counterpart. Interestingly, CD8+ T cells of donor C did not respond when primed, restimulated, and assayed with the unmodified peptide MAGE-A2250 (Table III), suggesting that the modified epitope is more potent in priming T cells. Once primed, these T cells can mount good responses to the natural variant as well (Fig. 3A) although the peptide concentrations required may be higher as shown for cytolysis of MAGE-A2116-pulsed T2 target cells by T cells primed with the corresponding anchor position-optimized variant of the epitope (Fig. 3B). The concentration of the natural epitope required for half-maximal cytolysis is 150-times higher than what is required in the case of the modified peptide. Similar results were obtained for TRP-2185 (data not shown). In a more stringent setting, selected T cell lines raised against the modified versions of these three epitopes by repeated stimulation in vitro were tested in chromium release assays with both variants (Fig. 4, A–C). In the effector-to-target titration, the response curves against both epitope variants are congruent indicating that, as far detectable by these assays, all T cells primed with the modified epitopes responded to the natural counterpart as well. Thus, both sets of data indicate that the natural and the modified epitopes are recognized by the same T cells. Because the cytotoxicity assays were done with the T2 cells, which share only the HLA-A*0201 with the donors of the T cells, these results also prove that the newly identified epitopes are indeed presented by this HLA allomorph.
Processing and presentation of the noncanonical epitopes by melanoma cells
To confirm that the predicted epitopes are naturally processed and presented by the tumor cells, we tested the above cytolytic CD8+ T cell lines from healthy donors that had been primed with the peptides TRP-2185mod, MAGE-A2116mod, or MAGE-A2250mod for their capacity to lyse human melanoma cells that express the target Ags together with or without HLA-A*0201 (Fig. 4, D–F). The tumor cell lines used were SK-mel 37, which expressed TRP-2 and MAGE-A2 as well as HLA-A*0201, and LRD-mel, which expresses both Ags but not HLA-A*0201. The TRP-2-specific CTL lyse SK-mel 37 indicating that the natural counterpart of the priming peptide is presented by these cells. This lysis is completely inhibited with the HLA A2-specific mAb BB7.2. The HLA restriction of these CTL is confirmed by the fact that LRD-mel cells that lack HLA-A*0201 but express the Ag are not lysed. Likewise, CTL raised against MAGE-A2116mod lysed Sk-mel 37 tumor cells but not the HLA A*0201-negative LRD-mel cells. Again this cell lysis was inhibited by the HLA A2-specific Ab BB7.2. Also, MAGE-A2250mod-specific CTL lysed Sk-mel 37 but not or only weakly the LRD-mel cell line, and the lysis of Sk-mel 37 is inhibited with the HLA A2-specific mAb BB7.2. To extend these analyses, the MAGE A2116mod-primed CTL line was tested at a fixed E:T ratio against additional melanoma cell lines. Sk-mel 37, Sk-mel 24, and Malme-3M express MAGE-A2 and HLA-A*0201 and were lysed. This lysis was inhibited by the HLA A*0201-specific mAb BB7.2. In contrast, the cell lines HRR-mel and LRD-mel that express the Ag but not HLA-A*0201, and AEO-mel that expresses HLA-A*0201 but not the Ag, were not lysed (Fig. 5A). Also, a cell line with specificity for TRP-2185mod generated from patient REN recognized naturally processed Ag presented by the melanoma cells Sk-mel 24, Sk-mel 37, and Malme-3M but not the TRP-2-positive, HLA-A*0201-negative cells SK-mel 28, HRR-mel, and IEL-mel (Fig. 5B). For both cell lines, specificity for their cognate epitope and its modified variant was demonstrated with peptide-pulsed T2 cells (Fig. 5). This series of experiments proves that, first, the three predicted epitopes TRP-2185, MAGE-A2116, or MAGE-A2250 are indeed naturally generated and presented by the tumor cells; second, the recognition of the epitopes by the CTL is HLA A*0201-restricted; and third, again CTL primed with the modified epitopes recognize their natural counterpart.
Discussion
The identification of T cell epitopes has been immensely advanced by the introduction of prediction algorithms for the identification of potential HLA ligands in the sequences of TAAs (37, 38, 39). The vast majority of T cell epitopes known to date were identified by bioinformatic prediction followed by experimental validation (31). Despite this success, the available algorithms are based on two critical assumptions that restrict the output of the computation. First, it is assumed that T cell epitopes need to bind strongly to the HLA molecules for efficient induction of T cells. Second, the contributions of the individual amino acids in the epitopes sequences to the strength of binding to the HLA molecules are treated as being independent from the sequence context. However, experimental data are published that contradict both of these assumptions (20, 21, 40). With the work presented in this study, we attempted to develop and test bioinformatic algorithms that take the sequence context within the peptide into consideration and can predict epitopes that do not conform to the above binding requirements. All seven noncanonical epitopes predicted with the ANN are indeed T cell epitopes. Five can trigger strong T cell responses in healthy donors. For four of the epitopes, high frequencies could be detected in melanoma patients, which are comparable to the T cell responses found in patients against the known canonical epitopes of the same Ags. Therefore, the corresponding epitopes may be classified as among the immunodominant epitopes of these Ags. T cells with specificity for the new epitopes can recognize and lyse melanoma cells showing that they are generated and presented naturally. Six of the seven T cell epitopes would not have been predicted by the other algorithms because they either lack the canonical HLA allele-specific epitope motifs or the calculated half-life times of their complexes with the HLA molecule is too short. In fact, HLA stabilization experiments with the TAP-deficient cell line T2 proved that these peptides are poor HLA ligands. Despite this low binding capacity, these new epitopes are capable of efficiently stimulating T cells as shown for cells from healthy donors after in vitro priming as well as for T cells from cancer patients, which respond to the peptides directly ex vivo. These results suggest that probably many potent T cell epitopes have been missed by the current approaches to T cell epitope determination even in such thoroughly studied TAAs as MAGE and gp100. They also testify to the power of sequence context-sensitive algorithms for epitope prediction, which are bound to substantially extend our knowledge of tumor antigenicity.
Improved suboptimal T cell epitopes generated by introducing the canonical anchor amino acids into the sequences have been proposed for a few epitopes, one of which, the MART-1/Melan-A26 epitope (41) has already been tested in clinical therapeutic vaccination trials with melanoma patients (42). However, this use of variant T cell epitopes with optimized HLA-binding properties has been controversial, because it was shown for some that T cells raised against the modified peptides do not efficiently recognize their natural counterparts as presented by the tumor cells and that fewer T cells in the peripheral blood of patients respond to natural than to the variant epitope (43, 44). However, in the examples presented with this report, all T cell lines induced with the optimized epitopes recognize their natural counterpart, also when presented naturally by the tumor cells. This outcome concords with the reports by several other investigators (45, 46). It obviously needs to be tested for every case whether anchor position-optimized epitopes can efficiently substitute for weak HLA binders and what modifications best guarantees efficient induction of antitumor T cell responses. Naturally presented weak HLA-binding epitopes may still induce T cells efficiently when their continuous production by the tumor cell translates into a sufficiently high steady-state representation at the cell surface. However, for use as vaccine Ag, it may be beneficial if the peptides bind fast and stably to the HLA molecules so to escape proteolytic degradation. The vaccination effect of suboptimal T cell epitopes may depend on improving their HLA-binding properties.
In summary, we demonstrate in this study that improved bioinformatics with ANN that predict potential T cell epitopes on the basis of considering the sequence context of the amino acids leads indeed to the identification of new epitopes that are not necessarily strong HLA binders, but nonetheless, are efficient inducers of T cell responses. These new tools will advance our understanding of tumor antigenicity and tumor immunity, enable us to monitor tumor-specific immune responses more comprehensively, and help to design improved cancer vaccines.
Acknowledgments
We thank Drs. Uwe Trefzer, Raoul Hasert, and Jan Gutermuth for providing the clinical material used in this study. The use of the human material for this study was reviewed and approved by the institutional ethics committee of the Charité (25.05.01).
Footnotes
The costs of publication of this article were defrayed in part by the payment of page charges. Th is article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
1 This work was supported by Deutsche Forschungsgemeinschaft Grants KFO 0050, FOR 299/2-1,2, and KL 427/11-1,2; CallistoGen Grant ANN-TAA01; and Deutsche Krebshilfe Grant 10-1898-Sp1.
2 A.B. and F.O.L. have contributed equally to the work.
3 Current address: IMS Health GmbH and Co. OHG, Hahn stra?e 30-32, 60528 Frankfurt/Main, Germany.
4 Address correspondence and reprint requests to Dr. Peter Walden, Department of Dermatology, Charité Campus Mitte, Clinical Research Group Tumor Immunology, Charité-Universit?tsmedizin Berlin, Schumannstrasse 20/21, 10117 Berlin, Germany. E-mail address: peter.walden{at}charite.de
5 Abbreviations used in this paper: TAA, tumor-associated Ag; ANN, artificial neural network; ICC, intracellular cytokine; TATE, tumor-associated T cell epitope; TRP-2, tyrosinase-related protein 2; RT, room temperature; BIMAS, BioInformatics and Molecular Analysis Section.
Received for publication September 20, 2004. Accepted for publication March 17, 2005.
References
Parmiani, G., C. Castelli, P. Dalerba, R. Mortarini, L. Rivoltini, F. M. Marincola, A. Anichini. 2002. Cancer immunotherapy with peptide-based vaccines: what have we achieved? where are we going?. J. Natl. Cancer Inst. 94: 805-818.
Rosenberg, S. A.. 2001. Progress in human tumour immunology and immunotherapy. Nature 411: 380-384.
Waldmann, T. A.. 2003. Immunotherapy: past, present and future. Nat. Med. 9: 269-277.
Romero, P., J. C. Cerottini, D. E. Speiser. 2004. Monitoring tumor antigen specific T-cell responses in cancer patients and phase I clinical trials of peptide-based vaccination. Cancer Immunol. Immunother. 53: 249-255.
Marchand, M., P. Weynants, E. Rankin, F. Arienti, F. Belli, G. Parmiani, N. Cascinelli, A. Bourlond, R. Vanwijck, Y. Humblet, et al 1995. Tumor regression responses in melanoma patients treated with a peptide encoded by gene MAGE-3. Int. J. Cancer 63: 883-885.
Jaeger, E., H. Bernhard, P. Romero, M. Ringhoffer, M. Arand, J. Karbach, C. Ilsemann, M. Hagedorn, A. Knuth. 1996. Generation of cytotoxic T-cell responses with synthetic melanoma-associated peptides in vivo: implications for tumor vaccines with melanoma-associated antigens. Int. J. Cancer 66: 162-169.
Marshall, J. L., R. J. Hoyer, M. A. Toomey, K. Faraguna, P. Chang, E. Richmond, J. E. Pedicano, E. Gehan, R. A. Peck, P. Arlen, et al 2000. Phase I study in advanced cancer patients of a diversified prime-and-boost vaccination protocol using recombinant vaccinia virus and recombinant nonreplicating avipox virus to elicit anti-carcinoembryonic antigen immune responses. J. Clin. Oncol. 18: 3964-3973.(Anne Bredenbeck2,*,, Flor)
The identification of tumor-associated T cell epitopes has contributed significantly to the understanding of the interrelationship of tumor and immune system and is instrumental in the development of therapeutic vaccines for the treatment of cancer. Most of the known epitopes have been identified with prediction algorithms that compute the potential capacity of a peptide to bind to HLA class I molecules. However, naturally expressed T cell epitopes need not necessarily be strong HLA binders. To overcome this limitation of the available prediction algorithms we established a strategy for the identification of T cell epitopes that include suboptimal HLA binders. To this end, an artificial neural network was developed that predicts HLA-binding peptides in protein sequences by taking the entire sequence context into consideration rather than computing the sum of the contribution of the individual amino acids. Using this algorithm, we predicted seven HLA A*0201-restricted potential T cell epitopes from known melanoma-associated Ags that do not conform to the canonical anchor motif for this HLA molecule. All seven epitopes were validated as T cell epitopes and three as naturally processed by melanoma tumor cells. T cells for four of the new epitopes were found at elevated frequencies in the peripheral blood of melanoma patients. Modification of the peptides to the canonical sequence motifs led to improved HLA binding and to improved capacity to stimulate T cells.
Introduction
The identification of tumor-associated Ags (TAA)5 and tumor-associated T cell epitopes (TATE) has led to new therapeutic vaccination strategies for the treatment of cancer (1, 2, 3) and is the basis for monitoring tumor-specific immune responses in patients (4). Since the mid-1990s, cancer vaccines of various designs such as synthetic peptides administered alone or together with IFAs or IL-2, or loaded onto autologous dendritic cells were tested in clinical trials (5, 6, 7, 8, 9). Although some trials have shown promising clinical responses, these were usually seen in no more than 20% of the patients. This low efficacy of therapeutic vaccination against cancer may, among other possible reasons, be caused by anergy, low functional avidity, or low frequency of the tumor-specific T cells (10, 11), or by a selection of Ag-loss variants of the tumor cells (12). It is now generally accepted that anti-tumor vaccines should include many different T cell epitopes to reduce the risk of immune evasion, and to address and activate as many tumor-specific T cells as possible. Consequently, the search for new TATE is greatly intensified. The most widely applied strategy for the identification of T cell epitopes, dubbed "reverse immunology," uses bioinformatic tools to predict peptides from the sequences of TAA that have the capacity to bind to the HLA molecules of interest. The predicted epitopes then need to be synthesized, tested in T cell assays, and validated as naturally processed and presented by the tumor cells (13, 14). The bioinformatics is based on HLA allele-specific sequence motifs that correspond to the amino acids that anchor the peptides in the peptide-binding grooves of the HLA molecules (15, 16, 17). The predictions use algorithms that are trained with the frequencies of the different amino acids at the different sequence positions of an epitope (18) or with binding data of known epitopes (19). As an example, HLA-A*0201, the most frequent HLA class I allomorph in Caucasian populations, binds preferentially nonapeptides with the aliphatic amino acids leucine, isoleucine, valine, or methionine at positions 2 and 9. The majority of the T cell epitopes known to date have been identified with such bioinformatic approaches. Notwithstanding, the available algorithms are very restrictive in that they select epitopes for their HLA-binding capacity and base these predictions predominantly on the individual input of the anchor amino acids. However, T cell epitopes need not be good HLA-binding peptides, as continuous production by the Ag-processing machinery may ensure sufficient representation at the cell surface for T cell stimulation. Moreover, it is expected and has been shown that, in HLA-restricted T cell epitopes, the entire sequence context and not only the amino acid occupancy at the individual sequence positions determines the HLA-binding and T cell-stimulating capacity of the peptides (20, 21). To overcome the limitations of the available bioinformatic tools, we developed a prediction algorithm that uses multilayer artificial neural networks (ANN) that, by their parallel data processing, compute the HLA-binding properties of peptides based on the entire sequence context and, thus, can account for the interdependences of the individual amino acids in an epitope. To test this algorithm we predicted HLA A*0201-restricted epitopes of the melanoma-associated TAAs gp100, MAGE-A1, MAGE-A2, p53, and tyrosinase-related protein 2 (TRP-2), and chose for further analyses seven of the predicted peptides that do not conform to the HLA allele-specific epitope sequence motifs for this HLA allomorph.
Materials and Methods
Intracellular cytokine (ICC) staining
ICC staining for epitope-specific CD8+ T cells in the peripheral blood of melanoma patients was described elsewhere (27). Briefly, PBMC were used after cryopreservation and incubated for 30 min at RT and 30 min at 37°C in DMEM (Invitrogen Life Technologies) with 0.1% BSA (Sigma-Aldrich) with the peptides at a concentration of 10 μg/ml. Then, equal volumes of 20 μg of BFA per milliliter of DMEM with 20% FCS was added, and the incubation continued for 5 h. The cells were pelletted, resuspended in PBS/1 mM EDTA, and incubated for 10 min at 37°C. After centrifugation and resuspension, lysis buffer (BD Biosciences) was added, the cells incubated for 3 min at RT, then washed and incubated with permeabilization buffer (BD Biosciences) for 10 min at RT in the dark. For flow cytometry, the cells were stained with anti-CD8-allophycocyanin, anti-CD3-PerCP, anti-CD69-PE, and anti-IFN--FITC (all BD Biosciences) and analyzed with a FACSCalibur flow-cytometer and CellQuest software (BD Biosciences).
Results
Prediction of noncanonical T cell epitopes
This study aimed at identifying with bioinformatic means HLA class I-restricted TATEs that bind only weakly to the corresponding HLA molecules and, therefore, are missed by other algorithms that are designed to predict peptides that bind strongly to HLA molecules and that conform to the HLA allele-specific epitope motifs as defined by the canonical amino acids listed in the SYFPEITHI database (Ref. 12 ; see introduction). To identify epitopes independent from these restrictive criteria we developed a strategy of in silico pattern recognition of HLA-binding T cell epitopes by ANN, which by their parallel mode of data processing can extract common features of complex amino acid sequence patterns (22, 23). Besides the specific design of the ANNs and the selection of input data sets, the predictions depend critically on the descriptors used to code the amino acids. We used the physicochemical-properties hydrophobicity, polarity, refractivity, and side chain volume as referenced (28, 29, 30). The ANNs were applied to MAGE-A1 and A2, gp100, p53, and TRP-2 to identify HLA A*0201-restricted epitopes. The output data included the known T cell epitopes gp100178 MLGTHTMEV, gp100619 RLMKQDFSV; TRP-2180 SVYDFFVWL, TRP-2288 SLDDYNHLV, TRP-2360 TLDSQVMSL; and MAGE-A2112 KMVELVHFL that feature canonical anchors as well as gp100154 KTWGQYWQV, gp100280 YLEPGPVTA, gp100209 ITDQVPFSV, gp100639 RLPRIFCSC; and MAGE-A1278 KVLEYVIKV that deviate from these motifs (31, 32, 33). The seven newly predicted epitopes together with variants with the canonical leucine at the anchor positions 2 and 9 were studied for their capacity to bind to HLA-A*0201 and to stimulate specific T cells responses.
Table I lists the 7 predicted epitopes and the variants with leucines at positions 2 and 9, together with the scores calculated by our ANN. The algorithms had been trained with peptide sequences classified as either HLA-A*0201 ligands or as negative for HLA A*0201-binding. Correspondingly, the output defines the test sequences as potential epitopes or as negative without grading, meaning that any number in Table I column ANN is a positive output. Thus, all 7 peptides and their variants were classified as potential epitopes despite differences in the scoring. The 14 peptides were also scored with the algorithms from the BioInformatics and Molecular Analysis Section (BIMAS) of the Center for Information Technology, National Institutes of Health, which is based on kinetic parameter of peptide-binding to the HLA molecules and predicts the half-life times of the respective HLA peptide complexes (19). Here, only the peptide TRP-2185 is identified as a potential epitope, none of the other 6 peptides are (Table I). In contrast, all of the variants peptides with the motif amino acid leucine in the anchor positions were scored highly. Consequently, TRP-2185 had been identified earlier with BIMAS (34), but none of the other peptides. Also the SYFPEITHI algorithm, which scores peptides according to the degree to which they conform to the HLA allele-specific epitope motifs indicates improved HLA-binding properties of the modified peptides (18). The SYFPEITHI scoring system values the amino acids in specific sequence positions that are conserved among different epitopes, and sums up these individual values to assess the capacity of the peptide to bind to the HLA molecule. Canonical amino acids in the dominant anchor positions get a value of 10. Because there are two dominant anchor positions that define the HLA allele-specific epitope motifs, strong binders should have a score of at least 20. A value below 20 indicates that at least one anchor position carries a suboptimal amino acid. Therefore, 20 is usually taken as the cut-off for T cell epitope prediction. By this criterion, none of the 7 peptides selected by our approach would qualify. However, the modified peptides do so as they were purposely adapted to the canonic anchor motifs (Table I). For an experimental assessment of the capacity of the 14 peptides to bind to HLA-A*0201, we performed HLA stabilization assays with the TAP-deficient cell line T2. Because of the TAP defect these cells lack the proper peptides for binding to and stabilizing of the native conformation of the HLA class I molecules. As a consequence, the cell surface expression levels of HLA are low but can increase when the cells are incubated with suitable peptides. The levels of HLA expression detected after peptide-pulsing can then be used as measure for the strength of binding as illustrated for the influenza matrix protein peptide MP58 in comparison with T2 cells without peptide (Fig. 1). Except for TRP-2185, HLA-A*0201 expression levels after incubation with the original peptides were unaltered when compared with the background without peptide. For TRP-2185 that already in its unmodified form scored highly in BIMAS and SYFPEITHI predictions, the HLA expression after incubation with the natural epitope was high above background but still topped by the introduction of leucine into the anchor positions. For MAGE-A1237 and gp100286, the modifications resulted in peptides that, in contrast to their natural counterparts, stabilized the HLA molecules indicating enhanced binding. Consistent with the predictions by the BIMAS algorithms, the modification of peptide p53347 did not produce a peptide with good HLA-binding properties. These results prove that six of the seven peptides predicted by our algorithms are weak HLA binders and would not be predicted as T cell epitopes by the other algorithms.
After in vitro priming and restimulation of CD8+ T cells from the peripheral blood of five healthy HLA A*0201-positive donors we could detect by ELISPOT assays specific responses against all seven epitopes. MAGE-A2116, MAGE-A1237, p5317, p53347, and TRP-2185 induced strong response (Table III). For MAGE-A2250 and gp100286 strong responses were seen only against the modified but not the natural epitopes. Two of the donors responded to six epitopes each, one against three and two against five epitopes. Taking the two sets of analyses with PBMC from melanoma patients and healthy donors together, all of the seven predicted noncanonical HLA-binding epitopes did induce specific T cell responses, six of them (MAGE-A2116, MAGE-A1237, gp100286, p5317, p53347, and TRP-2185) strong responses. Five of these (MAGE-A2116, MAGE-A1237, MAGE-A2250, gp100286, and TRP-2185) had already primed high frequencies of T cells in the patients. TRP-2185 had been identified before as an HLA A*0201-binding peptide but not yet as a T cell epitope (34). MAGE-A2116 was reported earlier as an HLA A3-restricted epitope (36). We show in this study for the first time that TRP-2185 is indeed a T cell epitope and that MAGE-A2116 is, in addition to HLA-A3, presented by HLA-A*0201. Six of the seven new HLA A*0201-restricted T cell epitopes were not identified by any of the other prediction algorithms for strong HLA-binding epitopes.
Comparison of the natural and the modified versions of the predicted epitopes
Several epitopes were recognized by the T cells in their natural as well as modified form. Therefore, we tested by ELISPOT whether T cells primed with the modified versions of the epitopes TRP-2185, MAGE-A2116, or MAGE-A2250 also respond to the natural variant of these peptides. In all cases, the CD8+ T cells primed with the modified epitope responded to both versions (Fig. 3A). In the case of TRP-2185, the responses to the natural epitope were higher, whereas for the other two, the T cells responded better to the priming peptide than to their natural counterpart. Interestingly, CD8+ T cells of donor C did not respond when primed, restimulated, and assayed with the unmodified peptide MAGE-A2250 (Table III), suggesting that the modified epitope is more potent in priming T cells. Once primed, these T cells can mount good responses to the natural variant as well (Fig. 3A) although the peptide concentrations required may be higher as shown for cytolysis of MAGE-A2116-pulsed T2 target cells by T cells primed with the corresponding anchor position-optimized variant of the epitope (Fig. 3B). The concentration of the natural epitope required for half-maximal cytolysis is 150-times higher than what is required in the case of the modified peptide. Similar results were obtained for TRP-2185 (data not shown). In a more stringent setting, selected T cell lines raised against the modified versions of these three epitopes by repeated stimulation in vitro were tested in chromium release assays with both variants (Fig. 4, A–C). In the effector-to-target titration, the response curves against both epitope variants are congruent indicating that, as far detectable by these assays, all T cells primed with the modified epitopes responded to the natural counterpart as well. Thus, both sets of data indicate that the natural and the modified epitopes are recognized by the same T cells. Because the cytotoxicity assays were done with the T2 cells, which share only the HLA-A*0201 with the donors of the T cells, these results also prove that the newly identified epitopes are indeed presented by this HLA allomorph.
Processing and presentation of the noncanonical epitopes by melanoma cells
To confirm that the predicted epitopes are naturally processed and presented by the tumor cells, we tested the above cytolytic CD8+ T cell lines from healthy donors that had been primed with the peptides TRP-2185mod, MAGE-A2116mod, or MAGE-A2250mod for their capacity to lyse human melanoma cells that express the target Ags together with or without HLA-A*0201 (Fig. 4, D–F). The tumor cell lines used were SK-mel 37, which expressed TRP-2 and MAGE-A2 as well as HLA-A*0201, and LRD-mel, which expresses both Ags but not HLA-A*0201. The TRP-2-specific CTL lyse SK-mel 37 indicating that the natural counterpart of the priming peptide is presented by these cells. This lysis is completely inhibited with the HLA A2-specific mAb BB7.2. The HLA restriction of these CTL is confirmed by the fact that LRD-mel cells that lack HLA-A*0201 but express the Ag are not lysed. Likewise, CTL raised against MAGE-A2116mod lysed Sk-mel 37 tumor cells but not the HLA A*0201-negative LRD-mel cells. Again this cell lysis was inhibited by the HLA A2-specific Ab BB7.2. Also, MAGE-A2250mod-specific CTL lysed Sk-mel 37 but not or only weakly the LRD-mel cell line, and the lysis of Sk-mel 37 is inhibited with the HLA A2-specific mAb BB7.2. To extend these analyses, the MAGE A2116mod-primed CTL line was tested at a fixed E:T ratio against additional melanoma cell lines. Sk-mel 37, Sk-mel 24, and Malme-3M express MAGE-A2 and HLA-A*0201 and were lysed. This lysis was inhibited by the HLA A*0201-specific mAb BB7.2. In contrast, the cell lines HRR-mel and LRD-mel that express the Ag but not HLA-A*0201, and AEO-mel that expresses HLA-A*0201 but not the Ag, were not lysed (Fig. 5A). Also, a cell line with specificity for TRP-2185mod generated from patient REN recognized naturally processed Ag presented by the melanoma cells Sk-mel 24, Sk-mel 37, and Malme-3M but not the TRP-2-positive, HLA-A*0201-negative cells SK-mel 28, HRR-mel, and IEL-mel (Fig. 5B). For both cell lines, specificity for their cognate epitope and its modified variant was demonstrated with peptide-pulsed T2 cells (Fig. 5). This series of experiments proves that, first, the three predicted epitopes TRP-2185, MAGE-A2116, or MAGE-A2250 are indeed naturally generated and presented by the tumor cells; second, the recognition of the epitopes by the CTL is HLA A*0201-restricted; and third, again CTL primed with the modified epitopes recognize their natural counterpart.
Discussion
The identification of T cell epitopes has been immensely advanced by the introduction of prediction algorithms for the identification of potential HLA ligands in the sequences of TAAs (37, 38, 39). The vast majority of T cell epitopes known to date were identified by bioinformatic prediction followed by experimental validation (31). Despite this success, the available algorithms are based on two critical assumptions that restrict the output of the computation. First, it is assumed that T cell epitopes need to bind strongly to the HLA molecules for efficient induction of T cells. Second, the contributions of the individual amino acids in the epitopes sequences to the strength of binding to the HLA molecules are treated as being independent from the sequence context. However, experimental data are published that contradict both of these assumptions (20, 21, 40). With the work presented in this study, we attempted to develop and test bioinformatic algorithms that take the sequence context within the peptide into consideration and can predict epitopes that do not conform to the above binding requirements. All seven noncanonical epitopes predicted with the ANN are indeed T cell epitopes. Five can trigger strong T cell responses in healthy donors. For four of the epitopes, high frequencies could be detected in melanoma patients, which are comparable to the T cell responses found in patients against the known canonical epitopes of the same Ags. Therefore, the corresponding epitopes may be classified as among the immunodominant epitopes of these Ags. T cells with specificity for the new epitopes can recognize and lyse melanoma cells showing that they are generated and presented naturally. Six of the seven T cell epitopes would not have been predicted by the other algorithms because they either lack the canonical HLA allele-specific epitope motifs or the calculated half-life times of their complexes with the HLA molecule is too short. In fact, HLA stabilization experiments with the TAP-deficient cell line T2 proved that these peptides are poor HLA ligands. Despite this low binding capacity, these new epitopes are capable of efficiently stimulating T cells as shown for cells from healthy donors after in vitro priming as well as for T cells from cancer patients, which respond to the peptides directly ex vivo. These results suggest that probably many potent T cell epitopes have been missed by the current approaches to T cell epitope determination even in such thoroughly studied TAAs as MAGE and gp100. They also testify to the power of sequence context-sensitive algorithms for epitope prediction, which are bound to substantially extend our knowledge of tumor antigenicity.
Improved suboptimal T cell epitopes generated by introducing the canonical anchor amino acids into the sequences have been proposed for a few epitopes, one of which, the MART-1/Melan-A26 epitope (41) has already been tested in clinical therapeutic vaccination trials with melanoma patients (42). However, this use of variant T cell epitopes with optimized HLA-binding properties has been controversial, because it was shown for some that T cells raised against the modified peptides do not efficiently recognize their natural counterparts as presented by the tumor cells and that fewer T cells in the peripheral blood of patients respond to natural than to the variant epitope (43, 44). However, in the examples presented with this report, all T cell lines induced with the optimized epitopes recognize their natural counterpart, also when presented naturally by the tumor cells. This outcome concords with the reports by several other investigators (45, 46). It obviously needs to be tested for every case whether anchor position-optimized epitopes can efficiently substitute for weak HLA binders and what modifications best guarantees efficient induction of antitumor T cell responses. Naturally presented weak HLA-binding epitopes may still induce T cells efficiently when their continuous production by the tumor cell translates into a sufficiently high steady-state representation at the cell surface. However, for use as vaccine Ag, it may be beneficial if the peptides bind fast and stably to the HLA molecules so to escape proteolytic degradation. The vaccination effect of suboptimal T cell epitopes may depend on improving their HLA-binding properties.
In summary, we demonstrate in this study that improved bioinformatics with ANN that predict potential T cell epitopes on the basis of considering the sequence context of the amino acids leads indeed to the identification of new epitopes that are not necessarily strong HLA binders, but nonetheless, are efficient inducers of T cell responses. These new tools will advance our understanding of tumor antigenicity and tumor immunity, enable us to monitor tumor-specific immune responses more comprehensively, and help to design improved cancer vaccines.
Acknowledgments
We thank Drs. Uwe Trefzer, Raoul Hasert, and Jan Gutermuth for providing the clinical material used in this study. The use of the human material for this study was reviewed and approved by the institutional ethics committee of the Charité (25.05.01).
Footnotes
The costs of publication of this article were defrayed in part by the payment of page charges. Th is article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
1 This work was supported by Deutsche Forschungsgemeinschaft Grants KFO 0050, FOR 299/2-1,2, and KL 427/11-1,2; CallistoGen Grant ANN-TAA01; and Deutsche Krebshilfe Grant 10-1898-Sp1.
2 A.B. and F.O.L. have contributed equally to the work.
3 Current address: IMS Health GmbH and Co. OHG, Hahn stra?e 30-32, 60528 Frankfurt/Main, Germany.
4 Address correspondence and reprint requests to Dr. Peter Walden, Department of Dermatology, Charité Campus Mitte, Clinical Research Group Tumor Immunology, Charité-Universit?tsmedizin Berlin, Schumannstrasse 20/21, 10117 Berlin, Germany. E-mail address: peter.walden{at}charite.de
5 Abbreviations used in this paper: TAA, tumor-associated Ag; ANN, artificial neural network; ICC, intracellular cytokine; TATE, tumor-associated T cell epitope; TRP-2, tyrosinase-related protein 2; RT, room temperature; BIMAS, BioInformatics and Molecular Analysis Section.
Received for publication September 20, 2004. Accepted for publication March 17, 2005.
References
Parmiani, G., C. Castelli, P. Dalerba, R. Mortarini, L. Rivoltini, F. M. Marincola, A. Anichini. 2002. Cancer immunotherapy with peptide-based vaccines: what have we achieved? where are we going?. J. Natl. Cancer Inst. 94: 805-818.
Rosenberg, S. A.. 2001. Progress in human tumour immunology and immunotherapy. Nature 411: 380-384.
Waldmann, T. A.. 2003. Immunotherapy: past, present and future. Nat. Med. 9: 269-277.
Romero, P., J. C. Cerottini, D. E. Speiser. 2004. Monitoring tumor antigen specific T-cell responses in cancer patients and phase I clinical trials of peptide-based vaccination. Cancer Immunol. Immunother. 53: 249-255.
Marchand, M., P. Weynants, E. Rankin, F. Arienti, F. Belli, G. Parmiani, N. Cascinelli, A. Bourlond, R. Vanwijck, Y. Humblet, et al 1995. Tumor regression responses in melanoma patients treated with a peptide encoded by gene MAGE-3. Int. J. Cancer 63: 883-885.
Jaeger, E., H. Bernhard, P. Romero, M. Ringhoffer, M. Arand, J. Karbach, C. Ilsemann, M. Hagedorn, A. Knuth. 1996. Generation of cytotoxic T-cell responses with synthetic melanoma-associated peptides in vivo: implications for tumor vaccines with melanoma-associated antigens. Int. J. Cancer 66: 162-169.
Marshall, J. L., R. J. Hoyer, M. A. Toomey, K. Faraguna, P. Chang, E. Richmond, J. E. Pedicano, E. Gehan, R. A. Peck, P. Arlen, et al 2000. Phase I study in advanced cancer patients of a diversified prime-and-boost vaccination protocol using recombinant vaccinia virus and recombinant nonreplicating avipox virus to elicit anti-carcinoembryonic antigen immune responses. J. Clin. Oncol. 18: 3964-3973.(Anne Bredenbeck2,*,, Flor)