Peptide-MHC class I stability is a better predictor than peptide affinity of CTL immunogenicity


CorrespondenceProf. Søren Buus, Laboratory of Experimental Immunology, Panum 18.3.12, Blegdamsvej 3B, DK-2200 Copenhagen N, Denmark

Fax: +45-35327696


Additional correspondence Dr. Mikkel Harndahl, Laboratory of Experimental Immunology, Panum 18.3.12, Blegdamsvej 3B, DK-2200 Copenhagen N, Denmark.



Efficient presentation of peptide-MHC class I (pMHC-I) complexes to immune T cells should benefit from a stable peptide-MHC-I interaction. However, it has been difficult to distinguish stability from other requirements for MHC-I binding, for example, affinity. We have recently established a high-throughput assay for pMHC-I stability. Here, we have generated a large database containing stability measurements of pMHC-I complexes, and re-examined a previously reported unbiased analysis of the relative contributions of antigen processing and presentation in defining cytotoxic T lymphocyte (CTL) immunogenicity [Assarsson et al., J. Immunol. 2007. 178: 7890–7901]. Using an affinity-balanced approach, we demonstrated that immunogenic peptides tend to be more stably bound to MHC-I molecules compared with nonimmunogenic peptides. We also developed a bioinformatics method to predict pMHC-I stability, which suggested that 30% of the nonimmunogenic binders hitherto classified as “holes in the T-cell repertoire” can be explained as being unstably bound to MHC-I. Finally, we suggest that nonoptimal anchor residues in position 2 of the peptide are particularly prone to cause unstable interactions with MHC-I. We conclude that the availability of accurate predictors of pMHC-I stability might be helpful in the elucidation of MHC-I restricted antigen presentation, and might be instrumental in future search strategies for MHC-I epitopes.


Major histocompatibility complex class I (MHC-I) plays a pivotal role in the generation of specific immune responses mediated by cytotoxic T lymphocytes (CTLs). MHC-I molecules sample peptides derived from intracellular proteins, translocate them to the cell surface, and display them to CTLs, allowing immune scrutiny of the ongoing intracellular metabolism leading to the detection of the presence of any intracellular pathogens. To fulfill this crucial antigen presenting function, MHC-I molecules must be endowed with the ability to retain bound peptides at the cell surface while waiting for the arrival of rare circulating CTL clones of the appropriate specificity. Sustained presentation at the cell surface and induction of specific immune T-cell responses therefore requires some degree of pMHC-I stability. Indeed, it has been claimed that stability, rather than affinity, of pMHC-I complexes is the better correlate of immunogenicity and immunodominance [[1-5]]. Experimentally, however, affinity remains the most frequently established correlate of immunogenicity. Thus, when Assarsson et al. [[6]] recently conducted a quantitative analysis of the variables affecting the repertoire of T-cell specificities recognized after vaccinia virus infection, they found that the vast majority of epitopes (85%) bound their restricting allele with an affinity of 500 nM or better, and most (75%) bound with an affinity of 100 nM or better. Investigating the stability of pMHC-I complexes for a small sample of immunogenic and nonimmunogenic peptides, they found a suggestive, but not statistically significant, trend for off-rates and immunodominance being correlated. The authors concluded that “in our hands, peptide stability did not correlate significantly better with immunodominance than did equilibrium binding measurements”.

One reason why pMHC stability has not been addressed more extensively undoubtedly relates to the cumbersome and/or low-throughput nature of current biochemical methods used to measure the dissociation of pMHC complexes [[6-12]]. A particularly interesting dissociation assay developed by Parker et al. [[13]] uses the dissociation of radiolabeled beta-2-microglobulin (β2m) as a convenient measure of peptide dissociation; however, the capacity of this assay is limited. We have recently generated a high-throughput, homogenous version of this assay, based upon a scintillation proximity principle allowing online, real-time monitoring of the dissociation of 125I-labeled β2m from recombinant MHC-I heavy chains [[14]]. Here, we have used this assay to address the stability of immunogenic and nonimmunogenic pMHC-I complexes. Using panels of affinity-balanced peptides, we could demonstrate that the stabilities of pMHC-I complexes involving known T-cell epitopes are significantly more stable than pMHC-I complexes involving peptides of similar-binding affinity that are not known to be immunogenic. Our results also suggest that HLA-A*02:01-binding peptides become destabilized if the P2 anchor residue, and to a lesser extend the P9 anchor residue, are not optimal; and that anchor optimization increases both affinity and stability. In conclusion, our results suggest that some peptides, despite exhibiting high-affinity binding to HLA class I molecules, may fail to become immunogenic because they fail to form stable complexes with HLA class I molecules.


Determining affinity and stability of peptide-MHC-I complexes

We used high-throughput homogenous biochemical assays to measure the affinity and stability of pMHC-I complexes [[14, 15]]. To generate pMHC-I complexes, biotinylated MHC-I heavy chain molecules were diluted more than 100-fold into a folding buffer containing β2m and peptide; and incubated to reach steady-state pMHC-I complex formation. All in vitro biochemical peptide-MHC-I affinity measurements (and all complex formation for subsequent dissociation experiments) were done at 18°C to avoid the confounding loss of complexes due to temperature instability [[16]]. In contrast, the dissociation phase of dissociation experiments was conducted at 37°C.

To measure the affinity of peptide-MHC-I interactions, dose–response experiments were done in order to determine the peptide concentration (EC50) resulting in half-saturation of folded pMHC-I complexes (Fig. 1A). A homogenous luminescence oxygen channeling immunoassay (LOCI) was used to measure the resulting formation of folded pMHC-I complexes [[15]]. Under conditions of limited receptor concentration ([MHC-I HC] ≤ KD), the EC50 is a reasonable approximation of the equilibrium dissociation constant, KD.

Figure 1.

Affinity and stability comparison of peptide-MHC-I interactions of two different peptides. (A) Affinity of a T-cell epitope (FLTSVINRV, •) and a nonimmunogenic peptide (NQNDNEETV, ○) was determined by dose–response experiments. Formation of folded MHC-I was measured with a homogenous luminescent oxygen channeling immunoassay (LOCI) on an EnVision multilabel reader [[15]]. (B) Dissociation of a T-cell epitope (FLTSVINRV, •) and a nonimmunogenic peptide (NQNDNEETV, ○) at 37°C was measured with a scintillation proximity assay (SPA) using radiolabeled β2m and streptavidin-coated scintillation microplates (Flashplates) [[14]]. The scintillation signal was measured repeatedly in a TopCount microplate scintillation counter and the stability was determined from dissociation curves. Results are representative of three independent experiments.

To measure the rate of peptide dissociation, we exploited an observation made initially by Parker et al. [[13]] showing that dissociation of 125I-labeled β2m is an accurate measurement of peptide dissociation. We recently showed that pMHC-I dissociation can conveniently be monitored in real time using a scintillation proximity assay (SPA) [[14]]. To this end, pMHC-I complexes were generated under conditions that led to optimal incorporation of 125I-labeled β2m. After steady-state had been reached, dissociation measurements were initiated by adding an excess of unlabeled β2m thereby blocking any reassociation of dissociated 125I-labeled β2m, and at the same time the temperature of the reaction was increased to 37°C. This setup allowed a robust determination of the rate of dissociation, kd (Fig. 1B).

Examining a large panel of peptide-HLA-A*02:01 complexes for affinity and stability

Although affinity and stability are directly related, a high-affinity peptide-MHC-I interaction does not necessarily translate into a high-stability interaction (Fig. 1). Even though the two peptides in this particular example, both had the same affinity of 7 nM to HLA-A*02:01 (Fig. 1A), the stability of the two pMHC-I complexes varied considerably: one being very stable with a half-life of about 22 h, and the other being quite unstable with a half-life of only about 1 h (Fig. 1B). To examine the relationship between affinity and stability in more detail, we analyzed a large panel of available HLA-A*02:01-binding peptides. During the past 5–10 years, we have collected more than 10,000 peptides according to different strategies: some have been selected, because they represent known T-cell epitopes (as entered into the SYFPEITHI and/or into the IEDB databases); some have been predicted to be MHC-I binders in various T-cell epitope discovery projects [[17, 18]]; some have been predicted to rationally populate the MHC-I-binding space [[19, 20]]; and some have been found by screening randomly selected peptides. Aiming at including all peptides that might have an immunogenic potential and erring at the lower side of the conventionally accepted affinity threshold for being immunogenic of 500 nM, we selected all peptides with a measured HLA-A*02:01-binding affinity stronger than about 1000 nM. A total of 739 peptides were available for this comparative analysis of stability versus affinity, which is depicted in a log(stability) versus log(affinity) plot (Fig. 2A and B). A total of 107 were known T-cell epitopes or natural ligands (Fig. 2A), 632 were not known to be T-cell epitopes (Fig. 2B). Both stable and unstable interactions could be found throughout this intermediate to high-binding range. Comparing the known immunogenic peptides to those that are not known to be immunogenic, immunogenicity was found to correlate significantly with high affinity and even more significantly with high stability (p = 0.017 and p = 0.0004, respectively, unpaired two-tailed Students t-test), (Fig. 2C and D).

Figure 2.

Relationship between affinity and stability of immunogenic peptides and peptides with unknown immunogenicity to HLA-A*02:01. (A) Correlation between affinity and stability of 107 known T-cell and natural (immunogenic) epitopes restricted to HLA-A*02:01 as measured by dose–response experiments using an LOCI [[15]] and an SPA assay [[14]], respectively. (B) Correlation between affinity and stability of 632 HLA-A*02:01-binding peptides with unknown immunogenic classification. (C) Comparison of the affinity of the immunogenic peptides presented in (A) and the affinity of the unknown peptides presented in (B). (D) Comparison of the stability of the immunogenic peptides presented in (A) and the stability of unknown peptides presented in (B). Data are shown as mean ± SEM. p-values were obtained using unpaired, two-tailed t-test.

Comparing stability of affinity-balanced epitopes of known vs. unknown immunogenicity

To address whether peptide-MHC-I stability is a better discriminator of immunogenicity than affinity, we paired peptides of known immunogenicity (extracted from the IEDB or SYFPEITHI databases) with peptides of equal affinity, but of unknown immunogenicity. Four HLA alleles were included in this study: A*01:01 (n = 17, average affinity 126 nM), A*02:01 (n = 42, average affinity 8 nM), B*07:02 (n = 9, average affinity 128 nM), and B*35:01 (n = 9, average affinity 125 nM). The stabilities between the immunogenic and unknown groups were compared (Fig. 3). For three of the four alleles (A*01:01, A*02:01, and B*35:01), we found that the immunogenic group was significantly more stable than the unknown group (p < 0.0001, p < 0.0001, and p = 0.0033, respectively, paired two-tailed Student's t-test), but for one allele, B*07:02, there was no significant difference in stability between the two groups (p = 0.22, paired two-tailed Student's t-test). This suggests that stability in general is a better indicator of immunogenicity than affinity is.

Figure 3.

Immunogenic peptides are more stably bound than peptides with unknown immunogenicity. Comparison of the stability of affinity balanced pairs, one immunogenic (●), and one of unknown immunogenicity (○), for four different HLA molecules. A*01:01 (n = 17), A*02:01 (n = 42), B*07:02 (n = 9), and B*35:01 (n = 9) as measured using SPA [[14]]. Data are shown as ± SD. p-values were obtained using unpaired, two-tailed t-test.

An unbiased analysis of the stability of immunogenic vs non-immunogenic binders

The above comparison of immunogenic peptides and peptides of unknown immunogenicity is potentially flawed. First, these peptides have been selected for purposes other than the present study and do not necessarily represent a random, representative and unbiased sample of the peptide space. Second, the data on these peptides are not particularly homogenous, since the database entries on immunogenicity are the result of the work over several decades by many different scientists using many different techniques. Third, the data might have be skewed due to the frequent use of predictions based on more or less complicated MHC-I-binding motifs, which may have led to an oversampling of peptides carrying perfect motif matches resulting in a likely overrepresentation of high-affinity and -stability binders. Fourth, the data are not error free. The immunogenic peptide sequences identified by synthesis and functional analysis do not necessarily represent the final stimulatory moieties (as first noted by Ploegh and colleagues [[21]]). Also, in most cases it has not been examined whether the peptide sequences used here as control peptides are truly nonimmunogenic (albeit the frequency of random peptides being immunogenic a priori is low [[22]]). Thus, one should be cautious when interpreting the data obtained with this panel of peptides.

To circumvent the above problem and reliably evaluate how affinity and/or stability correlate with immunogenicity, one should ideally perform a systematic and unbiased analysis of all possible overlapping peptides from a model antigen or organism; however, the resources required would be prohibitive. As a work-round, we analyzed the stability of peptide-HLA-A*02:01 complexes reported in a recent study by Sette and colleagues on the T-cell specificities recognized after infecting HLA-A*02:01 transgenic mice with vaccinia virus [[6]]. This is one of the most comprehensive and careful studies of its kind: it used a very broad HLA-A*02:01 motif definition to capture an estimated 99.8% of all possible 9- and 10-mer binders from a large collection of proteins known to be targeted by CTLs; and it examined the immunogenicity of a representative sample of high-affinity binding peptides both following vaccinia infection as well as after peptide immunization. The authors identified 15 immunogenic 9-mer or 10-mer peptides that were recognized by HLA-A*02:01-restricted CTL in the context of a natural infection (“immunogenic binders”), and 40 nonimmunogenic 9-mer or 10-mer peptides that despite binding with high affinity to HLA-A*02:01 were neither recognized in the context of a natural infection, nor when immunized with peptide in incomplete Freunds adjuvant (“nonimmunogenic binders”). To answer the question of whether affinity or stability is the better correlate of immunogenicity, we extracted 12 affinity-balanced pairs each consisting of an “immunogenic binder” and a “nonimmunogenic binder” according to Sette and colleagues [6]. These peptides were synthesized and affinity and stability of their interactions with HLA-A*02:01 was measured. This representative analysis showed that “immunogenic binders” were significantly more stably bound to HLA-A*02:01 than “nonimmunogenic binders” (p = 0.0007, paired two-tailed Student's t-test) (Table 1, Fig. 4B), whereas no significant difference in affinity was observed between the two groups (Table 1, Fig. 4A). Note that one of the reported immunogenic peptides, RTLLGLILFV, in our hands was a low-affinity, low-stability-binding peptide. Upon closer inspection, the N-terminally truncated peptide, TLLGLILFV, appeared to be a likely HLA-A*02:01-binding peptide. This peptide was synthesized and found to be a high-stability (half-life 33 h) peptide. We would like to suggest that TLLGLILFV is the real HLA-A*02:01-restricted CTL epitope.

Table 1. Twenty-four peptides (12 classified as immunogenic and 12 classified as nonimmunogenic) were analyzed with respect to affinity, stability, and predicted affinity
ClassificationSequenceAssarsson affinity, nMHarndahl affinity, nMPredicted affinity, nMObserved stability (T½) h
Figure 4.

Stability is a better indicator of immunogenicity than affinity. (A) Comparison of the affinity of 12 immunogenic (●) and 12 nonimmunogenic (○) as measured using LOCI [[15]]. (B) Comparison of the stability of 12 immunogenic (●) and 12 nonimmunogenic (○) as measured using SPA [[14]]. (C) Correlation between affinity and stability of the two groups. Data are shown as ± SD. p-values were obtained using unpaired, two-tailed t-test.

Depicting this data in a log(stability) versus log(affinity) plot showed that the increased stability of peptide-HLA-A*02:01 complexes involving “immunogenic binders” (y = 0.65x − 5.1, R2 = 0.65) versus “nonimmunogenic binders” (y = 0.75x − 4.5, R2 = 0.53) was seen throughout the binding range KD < 100 nM (Fig. 4C).

Suboptimal anchor residues in the peptide causes instability of the peptide-MHC-I complex

When we inspected the 2 × 12 affinity-paired peptides (24 in total), we noted that 10 of 12 peptides with optimal amino acids residues in both anchor position 2 (LM) and C-terminal (VLI) had a half-life of more than 5 h, whereas nine of 12 peptides with a suboptimal amino acid residue (typically T or Q in position 2 or C-terminally) had a half-life of less than 5 h. At face value, this highly significant distribution (p = 0.014, Chi-square test with Yates correction) suggests that peptide-HLA-A*02:01 complexes are destabilized by just one of the anchor positions being occupied with a suboptimal amino acid. For the seven peptides with suboptimal anchor residues, we substituted the suboptimal anchor residue with an optimal residue (leucine or methionine in position 2 and valine in C-terminal), and repeated the stability experiment. In all seven cases, the stability was improved (in six of the seven peptides, stability was increased by seven to tenfold), and four of the seven previously unstable peptides achieved a half-life better than 5 h, see Table 2. Thus, there appear to be a subtle difference in the specificity of high-affinity peptides, which may tolerate a suboptimal amino acid residue in an anchor position, and the specificity of high-stability peptides, which seems to be less inclined to tolerate suboptimal amino acid residue in anchor positions (in particular not in position 2).

Table 2. Seven peptides with suboptimal anchors were substituted to having optimal anchors in both positions. The stability and affinity was measured and compared to the wild-type peptide
ClassificationSubstitutionSequenceT½ hoursFold increase in T½Affinity nMFold increase in affinity

Using bioinformatics predictors to analyze the requirements for affinity vs. stability

Using the data on the affinity and stability of the above 739 peptides interacting with HLA-A*02:01, we developed two artificial neural network (ANN) predictors; one predicting the affinity and one predicting the stability of the interaction. Although the data was limited compared with that of our other binding predictors, which are based on data sets with sizes up of 150,000 data points, these early generation predictors did successfully capture significant aspects of affinity (Pearsons's correlation coefficient [PCC] = 0.643 and AUC = 0.849, Fig. 5A) and stability (PCC = 0.680 and AUC = 0.906, Fig. 5B).

Figure 5.

Validation of affinity and stability predictors. (A and B) Correlation plots for 739 nonamerpeptides. Predictions were made using fivefold cross-validation for the individual networks for affinity and stability trained as described in Materials and methods. (A) Correlation between measured affinity (mA) and predicted affinity (pA). Affinity as measured using an LOCI assay [[15]]. (B) Correlation between measured stability (mS) and predicted stability (pS) as measured using SPA [[14]].

The availability of these predictors allowed us to address all the 9-mer peptides that were reported by Sette and colleagues as being high-affinity binders to HLA-A*02:01 (KD better than 100 nM): 12 “immunogens,” 6 “subdominant epitopes,” 29 “cryptic epitopes,” and 26 “nonimmunogens” [[6]]. Sette and colleagues define an immunogen is an epitope-specific T-cell response seen after infection; a subdominant epitope is an epitope-specific T-cell response seen after peptide immunization, that is capable of recognizing an infected target cell; a cryptic epitope is an epitope-specific T-cell response seen after peptide immunization that only recognizes a peptide pulsed target cell; and a nonimmunogen cannot induce an epitope-specific T-cell response, not even after peptide immunization. We noted that none of the dominant, subdominant, and cryptic epitopes had a predicted half-life of less than 1 h and we would like to suggest that this is a minimum stability threshold of immunogenic epitopes. At a half-life threshold of 1 h, eight of the 26 (31%) nonimmunogenic binders could be rejected (i.e. predicted to be low stability binders) without rejecting any of the immunogenic epitopes. At higher half-life thresholds, the stability predictor would begin to differentiate between dominant, subdominant, and cryptic epitopes suggesting a general order of stability: dominant > subdominant > cryptic epitopes > nonimmunogenic peptides (data not shown).

Next, we asked whether predicted stability is a better correlate of immunogenicity than predicted affinity is. A direct comparison showed predicted stability (as mentioned above rejecting eight of the 26 nonimmunogenic binders) as being a slightly better discriminator that predicted affinity (rejecting only four of the 26 at a conventional affinity threshold of 500 nM). This meager difference between stability and affinity is perhaps not that surprising since the two parameters are so closely related. To better differentiate between them, we implemented a baseline correction strategy. Comparing the transformed units of the affinity and stability ANN's, we could calculate a correlation between predicted binding and predicted stability (R2 = 0.72, data not shown), and then use this to perform an affinity-balancing baseline correction whereby the expected predicted stability of a peptide was estimated as a function of its predicted affinity. This analysis revealed that 11 out of 12 immunogenic peptides were predicted to bind with higher stabilities than estimated from their predicted affinities (Fig. 6). In contrast, the nonimmunogenic binders were evenly distributed around the corrected baseline (Fig. 6). The difference between the two groups of peptides was statistically highly significant (p < 0.001, unpaired, one-tailed t-test). Importantly, if we the reversed the baseline correction strategy and made it stability balancing; in effect asking whether affinity could provide a signal beyond stability suitable for differentiating between immunogenic and nonimmunogenic peptides, we did not find any significant difference between the two groups (p > 0.1, unpaired, one-tailed t-test). Thus, this bioinformatics-driven analysis suggested that predicted stability is a better discriminator of immunogenicity than predicted affinity is.

Figure 6.

Identifying immunogenic peptides using a bioinformatics-driven analysis. The predicted stability (in transformed units) is shown as a function of the predicted affinity (also in transformed units) for the immunogenic (12 solid circles) and nonimmunogenic (26 open circles) nonamer peptides from Sette and colleagues [[6]]. The insert gives the baseline corrected stability (S–S0) as a function of the predicted affinity.

Finally, addressing whether the two predictors identified any systematic differences in affinity motifs as compared with stability motifs, we randomly selected 500,000 natural 9-mer peptides, predicted their affinities and stabilities. Analyzing the upper 2% (10,000) predicted binders, we sorted them by predicted-binding affinity and split them in a pair-wise manner into two groups: a high-stability group and a low-stability group. In this way, the average predicted binding is equal between the two groups (p = 0.4, paired t-test). It was next calculated how large a fraction of the peptides in each group had preferred amino acids in each, or both, primary anchor position P2 and P9 where the preferred amino acids at P2 were L and M, and preferred amino acids at P9 were V, L, and I. The results of the analysis showed a significant reduction in the concurrent presence of both anchors in the group of low-stability peptides compared to high-stability peptide, and a corresponding increase in peptides missing optimal P2 anchor residues, but not in peptides missing optimal P9 anchor residues (Table 3). Thus, the ANN-driven analysis confirms the experimental findings that unstable binders tend to lack an optimal anchor residue in P2.

Table 3. The 2% (10,000) with highest predicted affinity out of 500,000 random peptides were paired on affinity. Each pair was split into a “stable” group (the peptide with the highest predicted stability) and an “unstable” group. The frequency of peptides having optimal anchors at position 2 and 9, denoted (+,+) was calculated together for each group. Likewise the frequency of peptides having only one optimal anchor, (+,−) or (−,+) or no optimal anchors (−,−)
P2(LM), P9(VLI)StableUnstablep-value
(+,+)30262155< 0.0001
(−,+)13132078< 0.0001
(−,−)66195< 0.0001


Many sequential processes are involved in both the generation and recognition of MHC-I-restricted CTL ligands. A picture of the sequence and relative contribution of these different processes in the generation of T-cell epitopes is emerging (as excellently reviewed in [[6, 22, 23]]), however, it is still incomplete and may still lack important undiscovered components [[6, 22, 23]]. Roughly, it has been estimated that one of 7–8 possible peptides are successfully generated by the processing machinery, that one in 50–200 processed peptides are successfully bound to MHC-I, and that one of two pMHC-I complexes are successfully matched by a corresponding specificity in the T-cell repertoire [[6, 22, 23]]. This would lead to an estimate of one in 500 to 2000 random peptides being immunogenic; however, Sette and colleagues have suggested that only one of 8000 random peptides are immunogenic and suggested that other, possibly unidentified, factors are important in defining immunogenicity [[6]]. Here, we will argue that the requirement for a stable MHC interaction is one of those “other” factors.

It is generally recognized that the requirement for binding and presentation by MHC-I molecules is by far the most selective event of antigen processing and presentation [[6, 22-24]]. When searching for CD8+ T-cell epitopes, an affinity better than 500 nM (termed a good binder) is commonly used as a threshold to select candidate immunogenic peptides [[25]]. Sette and colleagues recently estimated that “the vast majority of epitopes (85%) bound their restricting MHC-I with an affinity of 500 nM or better, and most (75%) bound with an affinity of 100 nM or better” [[6]]. Unfortunately, this criterion leads to the inclusion of many nonimmunogenic peptides (i.e. false positives). Others and we have observed that only some 10–20% of pathogen-derived peptides, which bind to MHC-I with an experimentally verified affinity of 500 nM, or better, are subsequently found to be immunogenic [[6, 25, 26]]. Testing the immunogenicity of all predicted immunogenic epitopes is currently a very slow, costly process, and any computational T-cell epitope discovery process would benefit from a better and more quantitative understanding of antigen processing and presentation.

It has been suggested that the stability of pMHC complex correlates with immunogenicity (both for MHC-I [[1, 27-32]], and for MHC-II [[2, 33]]); and it has even been suggested that stability correlates better with immunogenicity than affinity of peptide interaction with MHC-I [[34-37]] and MHC-II [[38]]. Common to all these reports is that the experimental data are limited to a few epitopes. Here, we have examined the stability of 739 peptides that bind to HLA-A*02:01 with an affinity of about 1000 nM or better. We found that the rate of dissociation at 37°C varied from a half-life of over 40 h to one of less than 0.1 h. To neutralize the effect of affinity, affinity-balanced pairs of known versus “not-known-to-be” immunogens restricted to different HLA alleles (A*01:01, A*02:01, B*07:02, and B*35:01) were extracted and analyzed biochemically. We found a highly significant difference in the stability of immunogens compared to “not-known-to-be” immunogens for three of the four HLA class I molecules examined. In parallel studies of the immunogenicity of HIV-derived epitopes restricted to B*57:02, B*57:03, B*58:01, B*07:02, B*42:01, and B*42:02, we have found that stability is a better discriminator of immunogenicity than affinity is (Kløverpris et al., manuscripts in preparation). Thus, the proposition that stability is a better indicator of immunogenicity can be extended to a wide range of HLA class I molecules.

We were, however, concerned that the underlying data set was not representative of an unbiased epitope discovery process, since many reported CTL epitopes have been discovered using simple rule-based predictions of high-affinity binding to MHC-I. We therefore turned to the unbiased study conducted by Sette and colleagues on the HLA-A*02:01-restricted T-cell epitopes found in vaccinia infections [[6]]. From this paper, we extracted an affinity-balanced set of 12 peptide pairs of immunogenic and nonimmunogenic binders, and tested both affinity and stability of their interaction with HLA-A*02:01. By and large, we confirmed the reported affinity data. Importantly, we found a highly significant correlation between high stability and immunogenicity (a half-life of 14 h for the immunogenic binders versus 3 h for the nonimmunogenic binders, p = 0.0007). Thus, this affinity-balanced reanalysis of the unbiased data reported by Sette and colleagues confirms that the stability of pMHC-I complexes contributes to the definition of immunogenicity, and that stability is a better indicator of immunogenicity than affinity is. This experimental analysis was subsequently corroborated by a bioinformatics-driven analysis.

Our data suggest that the description and relative contribution of antigen processing and presentation events needs to be redefined. Nonimmunogenic binders are usually considered to be the result of “holes in the T-cell repertoire.” However, a failure to achieve stable interaction with MHC may be considered an alternative mechanism of lacking immunogenicity, as a “hole in the stably bound MHC repertoire” mechanism, which would go unnoticed when solely addressing the affinity of peptide-MHC-I interactions. This may account for as much as 30% of these instances of lacking immunogenicity and it follows that a method to predict the stability of pMHC-I complexes might support computational CTL epitope discovery.

Finally, both our experimental and bioinformatics-driven analysis suggested that the lack of stability of HLA-A*02:01 binding occurred when the P2 anchor residue was not optimal. Although both threonine and glutamine can be found in P2 of high-affinity binding peptides, they lead to a seven to tenfold reduction in stability compared to the optimal leucine. Thus, it would appear that one anchor might be sufficient for binding, whereas two anchors might be needed to obtain stable MHC-I interaction. This is reminiscent of a previous suggestion that the different pockets of the MHC-II can be seen as interacting with peptide independently, and that destabilizing any of the pockets individually may lead to peptide dissociation [[39]]. Thus, pMHC-I stability studies should help elucidating how peptide bind and remain bound to MHC-I, and how MHC-I matures and eventually becomes a bona fide CTL target.

Materials and methods


Peptides were synthesized by Schafer-N (Copenhagen, Denmark) by Fmoc chemistry and HPLC purified to at least more than 80% (usually >95%), analyzed by HPLC and MS, and quantitated by weight.

MHC-I heavy chain proteins

Synthetic genes encoding MHC-I heavy chains were generated as previously described [[40, 41]]. Briefly, genes encoding MHC-I heavy chains truncated at position 275 (i.e. truncated before the membrane spanning domain), were fused C-terminally to a histidine-containing affinity tag (HAT) and a biotinylation signal peptide [[40, 42]]. The recombinant genes were expressed in the Escherichia coli expression host, BL21(DE3), harvested as inclusion bodies, extracted into a urea buffer and purified. The MHC-I heavy chain proteins were never exposed to reducing conditions. This allows purification of highly active preoxidized MHC-I heavy chains [[41]]. The proteins were identified by A280 absorbance and SDS-PAGE, and concentrations were determined by BCA assay (Pierce, Cat no. 23225). The degree of biotinylation (usually >95%) was determined by a gel-shift assay [[40]]. The preoxidized, denatured proteins were stored at −20°C in an 8 M urea buffer.

β2m protein

Native, recombinant human β2m was expressed and purified as previously described [[41, 42]]. Briefly, a HAT followed by an FXa restriction enzyme site was inserted N-terminally of a synthetic gene encoding the native, mature human β2m. The recombinant gene was expressed in the E. coli expression host, BL21(DE3), harvested as inclusion bodies, extracted into a urea buffer, folded by dilution and purified. The tagged β2m protein was digested for 48 h at room temperature with the FXa protease releasing intact natively folded β2m. The folded β2m was purified as previously described, and fractions containing β2m was identified by A280 UV absorbance and SDS-PAGE, and pooled. Protein concentrations were determined by BCA assay. The native β2m proteins were stored at −20°C.

Iodination of β2m

The recombinant β2m was radio-labeled with iodine (125I) using the chloramine-T procedure [[43]]. Twenty microgram of β2m was mixed with 1 mCi 125I and 5 μL chloramines-T (1 mg/mL) (Sigma, C9887, Sigma Alrich, Brondby, Denmark) for 1 min. The reaction was stopped by adding 5 μL metabisulfite (1 mg/mL) (Sigma). Unreacted iodine was removed by gel filtration chromatography using a 1 mL Sephadex G10 column equilibrated in PBS. Column fractions of 200 μL were tested for radioactivity and the labeled fractions were identified. The radioactivity was measured on a gamma counter (Packard Cobra 5010) and diluted to 25,000 cpm/μL in PBS containing 2% ethanol and 0.1% azide, and stored at 4°C.

Measuring peptide-MHC-I stability

The measurement of pMHC-I stability was done as recently described [[14]]. Briefly, recombinant, biotinylated MHC-I heavy chain molecules in 8 M urea were diluted 100-fold into PBS buffer containing radiolabeled β2m and peptide to initiate pMHC-I complex formation. The reaction was carried out in streptavidin coated scintillation 384 (or 96) well microplates (Flashplate® PLUS, Perkin Elmer, Boston, USA). The folding mixture, including 30 nM MHC-I heavy chain, trace amount of 125I-labeled β2m (approximately 1 nM, approximately 25,000 cpm) and saturating concentrations (10 μM) of the peptide, was allowed to reach steady-state (over-night incubation at 18°C), before unlabeled β2m was added to a final concentration of 1 μM to prevent reassociation of any dissociated labeled β2m, and the temperature was raised to 37°C. The dissociation was monitored by consecutive measurement of the scintillation microplate on a scintillation multiplate counter (TopCount NTX, Perkin Elmer), which was modified to operate at 37°C. Using a 384-well microtiter plate format, each well could be read approximately twice per hour.

Note that our biochemical stability assay compares favorably with the cellular-base stability assay reported by Wei et al. [[44]] where peptide-mediated stabilization of HLA-A*02:01 expression by the TAP-deficient cell line T2 was examined in the presence of brefeldin A, which prevented de novo HLA-A*02:01 expression and thus focused the assay on the stability of already expressed HLA-A*02:01 (data not shown).

Measuring peptide-MHC-I affinity

Affinity measurements of peptide interactions with MHC-I molecules were done using the AlphaScreen technology as previously described [[15]]. In brief, recombinant, biotinylated MHC-I heavy chains were diluted to a concentration of 2 nM in a mixture of 30 nM β2m and peptide, and allowed to fold for 48 h at 18°C. The pMHC-I complexes were detected using streptavidin donor beads and a conformation-dependent anti-HLA-I antibody, W6/32, conjugated to acceptor beads. The beads were added to a final concentration of 5 μg/mL and incubated over-night at 18°C. One hour prior to reading the plates, these were placed next to the reader to equilibrate to reader temperature. Detection was done on an EnVision multilabel reader (Perkin Elmer).

Data analysis

Association and dissociation curves were fitted using GraphPad Prism 5 (GraphPad, San Diego, CA, USA). Background subtracted dissociation data was fitted to a one-phase dissociation model:

display math

Construction of artificial neural network for HLA affinity and stability predictions

Conventional feed-forward artificial neural networks for stability and affinity predictions were constructed as earlier described by Nielsen et al. [[45]]. In brief, the networks have an input layer with 180 neurons, one hidden layer with ten neurons, and a single neuron output layer. The 180 neurons in the input layer encode the nine amino acids in the peptide sequence with each position represented by 20 neurons (one per amino acid type). The peptides were presented to the networks using sparse encoding, and the networks were trained in a fivefold cross/validation manner using the back-propagation procedure to update the weights in the network.

A total of 739 peptides with associated-binding affinities and binding stability values were used to train the neural networks. To boost the performance, the data were artificially enriched with 200 random natural negative peptides with assumed low affinity and stability [[46, 47]]. Binding affinity and stability values for the random negative peptides were set to 45,000 nM and 0.3 h, respectively, corresponding to transformed values (see below) of 0.01 for both affinity and stability. Binding affinity values were log-transformed using the relation 1−log50,000(EC50), and binding stability values were transformed using the relation (2–2/Th) in order to make both values fall in the range from 0 to 1. Here, EC50 is the binding affinity in nM units and Th is the half-life in hours. In each of the five cross-validations, fourth-fifth of the data were used to train a given network, and one-fifth was used to determine when to stop the training in order to avoid overfitting. Upon training, each prediction method (affinity and stability) thus consisted of an ensemble of five networks. When using the networks to predict binding of a query peptide, the prediction score is calculated as a simple average over the five networks in the given ensemble.


The authors thank Sara Pedersen for excellent technical assistance and Kenneth C. Parker for reviewing this manuscript. This work was supported by NIH grant HHSN272200900045C.

Conflict of interest

The authors declare no financial or commercial conflict of interest.


artificial neural network


luminescence oxygen channeling immunoassay


scintillation proximity assay