BIITE: A Tool to Determine HLA Class II Epitopes from T Cell ELISpot Data

Abstract

Activation of CD4+ T cells requires the recognition of peptides that are presented by HLA class II molecules and can be assessed experimentally using the ELISpot assay. However, even given an individual's HLA class II genotype, identifying which class II molecule is responsible for a positive ELISpot response to a given peptide is not trivial. The two main difficulties are the number of HLA class II molecules that can potentially be formed in a single individual (3-14) and the lack of clear peptide binding motifs for class II molecules. Here, we present a Bayesian framework to interpret ELISpot data (BIITE: Bayesian Immunogenicity Inference Tool for ELISpot); specifically BIITE identifies which HLA-II:peptide combination(s) are immunogenic based on cohort ELISpot data. We apply BIITE to two ELISpot datasets and explore the expected performance using simulations. We show this method can reach high accuracies, depending on the cohort size and the success rate of the ELISpot assay within the cohort.

Activation of CD4+ T cells requires the recognition of peptides that are presented by HLA class II molecules and can be assessed experimentally using the ELISpot assay. However, even given an individual's HLA class II genotype, identifying which class II molecule is responsible for a positive ELISpot response to a given peptide is not trivial. The two main difficulties are the number of HLA class II molecules that can potentially be formed in a single individual (3-14) and the lack of clear peptide binding motifs for class II molecules. Here, we present a Bayesian framework to interpret ELISpot data (BIITE: Bayesian Immunogenicity Inference Tool for ELISpot); specifically BIITE identifies which HLA-II:peptide combination(s) are immunogenic based on cohort ELISpot data. We apply BIITE to two ELISpot datasets and explore the expected performance using simulations. We show this method can reach high accuracies, depending on the cohort size and the success rate of the ELISpot assay within the cohort.

Author Summary

When studying the host immune response, a central question is: "which peptides elicit CD4+ T cell responses?" ELISpot assays are used to assess if subjects have responded to a given peptide. However, to determine which of the HLA-II molecules coded by the host HLA genotype is responsible for the reaction requires additional analysis. We present a Bayesian approach to solve this problem and have implemented it for use with the statistical language R under the BIITE moniker. Importantly, the aim of BIITE is to interpret experimental data, not to make in silico predictions. The method considers the immunogenicity of all HLA (in a cohort of patients) with respect to a given peptide simultaneously, in order to deal with linkage disequilibrium between genes of the HLA locus. Furthermore, users can enter additional information they might have (from literature or other experiments) in the form of prior information. The method is not exclusive to the HLA genes and can be used to attribute positive binary outcomes to any multi-allelic set of genes.

Citation: Boelen L, O'Neill PK, Quigley KJ, Reynolds CJ, Maillere B, Robinson JH, et al. (2016) BIITE: A Tool to Determine HLA Class II Epitopes from T Cell ELISpot Data. PLoS Comput Biol 12(3): e1004796. doi:10.1371/journal.pcbi.1004796

Editor: Bjoern Peters, La Jolla Institute for Allergy and Immunology, UNITED STATES

Received: September 2, 2015; Accepted: February 8, 2016; Published: March 8, 2016

Copyright: © 2016 Boelen et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: All software is available on http://github.com/liesb/BIITE and is

Fig 1. Outline of the algorithm.

We aim to describe the immunogenicity of a peptide:HLA combination as a number between 0 and 1. For each subject, the outcome of the ELISpot assay is known, together with a list of their HLA alleles. The user can add prior knowledge in the form of a prior distribution. In the example, the prior knowledge describes the belief that HLA class II allele A is not immunogenetic in combination with the peptide, while HLA class-II allele B is. The output of the algorithm consists of posterior marginal distributions of the values Ej. In the example, it is found that allele A is more immunogenic that allele B in combination with the peptide.

http://dx.doi.org/10.1371/journal.pcbi.1004796.g001

Validation on Burkholderia dataset

In order to test the BIITE algorithm, we used CD4+ T cell ELISpot data from a small cohort of exposed sero-positive individuals (N = 38). All overlapping 20-mer peptides from the BPSL2096 protein were tested (see Methods).

In Fig 2A, we present the marginal posterior distributions of the immunogenicity (solid lines) for all HLA molecules in combination with a single, representative peptide (BPSL2096, 1-20), calculated using a uniform prior. Especially for rare alleles (see S1 Table), the distributions are rather flat and uninformative. We also show the marginal posterior distributions, now assuming a non-uniform prior as described above (Fig 2A, dashed lines). The posterior distributions are now markedly more pronounced. This is of course driven by the prior assumptions, especially for the rare alleles. Nevertheless, there are divergences between the posterior and the prior. For example, for the representative peptide shown in Fig 2A both DRB1*01 (2 carriers) and DRB1*08 (4 carriers, one of which is homozygous), were assigned a positive prior (i.e. both were predicted to bind the peptide), yet they have different posteriors, reflecting the fact that both DRB1*01 carriers had positive ELISpot results, whilst there were two DRB1*08 carriers with a negative ELISpot.

Fig 2. Application of the algorithm to the Burkholderia data set, with (dashed lines) and without (solid lines) inclusion of prior information.

(A) Marginal posteriors of the peptide:HLA immunogenicities for BPSL2096 1-20 (Allele count in the cohort in brackets). (B) ROC curves were generated by comparing the mode of the posterior distributions against the transgenic mouse data set. For these curves, we have removed all posterior distributions whose Kullback-Leibler divergence with respect to the uniform distribution on the unit interval was smaller than that of the Beta(1,2) distribution.

http://dx.doi.org/10.1371/journal.pcbi.1004796.g002

We compared the posterior marginal distributions to the transgenic mouse model data (6 data points for each of the 17 peptides). To do so, one needs to collapse the posterior distribution to a decision: is the pHLA combination immunogenic or not, according to the model? For this analysis, we used the mode of the marginal posterior distribution as a summary statistic, and the decision rule was based on the mode being above a certain threshold. We then constructed a ROC curve by varying this threshold (Fig 2B). To interpret the strength of an algorithm, one compares it to the diagonal; this is the theoretical ROC curve for a random decision rule. The observed AUC is 71.3% when we use a uniform prior, and 65.4% when using a prior based on the predictions of NetMHCIIPan.

Validation on Pseudomonas dataset

We performed an additional analysis on a Pseudomonas dataset, comprised of N = 58 subjects. The AUCs we found were 68.8% (without predictive prior), 79.3% (with predictive prior) (see Fig 3).

Fig 3. ROC curves for the analysis of the Pseudomonas data set, with (dashed line) and without (solid line) inclusion of a predictive prior.

http://dx.doi.org/10.1371/journal.pcbi.1004796.g003

The subjects in the Pseudomonas study were divided into two groups, based on the number of positive sputum tests they had delivered. In this study, individuals with a high number of sputum cultures positive for PA had a lower immune responsiveness (lower number of peptide that elicited a T cell response) compared to individuals that were never sputum culture positive for PA, indicating that there were two different clinical groups with different peptide immunogenicities. Hence, we applied our algorithm to each of the groups separately, and predicted that the algorithm would perform poorly on the data from group 2 (who had very low response rates in their ELISpot assays). Indeed, for group 2 (subjects with at least one positive sputum sample), we found that BIITE was unsuccessful at resolving the HLA:epitope combinations found to be immunogenic in transgenic mice (AUC: 32% without prior; 69% with prior); this apparently low performance is not unexpected as these combinations were not immunogenic in humans. For group 1, BIITE reached AUCs of 69% (without prior) and 73% (with prior).

Determinants of performance

Cohort size.

To test the effect of the sample size of the ELISpot dataset, we modelled 17 in silico peptides based on a population of 200 subjects whose HLA class II haplotype was drawn (with replacement) from the Burkholderia cohort. We set peptide immunogenicity such that the odds ratio of having a positive ELISpot was similar to that in the Burkholderia dataset (S2 Fig). We then applied the algorithm to this new population and to subpopulations thereof. To validate the performance of the model, one needs to first decide which of the pHLA combinations are considered immunogenic. We did so by choosing a cut-off on the immunogenicity value Ej of each simulated pHLA combination; different cut-offs were considered (see Methods). For each of these cut-offs, we construct ROC curves and compute the AUC. Fig 4A shows the values of the AUC as a function of the cut-off on Ej and the population size. Unsurprisingly, performance of the model as measured by AUC increases with population size. Optimal AUCs were reached when the cut-off was taken at 0.4. Fig 4B shows the ROC curves associated with this cut-off. It is interesting to note that in our original Burkholderia dataset, where sample size was 38, we reached an AUC of 71%, which is comparable to the results found for sample sizes 30 and 50 here (72% and 75% respectively).

Fig 4. Simulated data based on the Burkholderia dataset.

(A) Evolution of the AUC with sample size, for different cut-offs on what is considered to be an immunogenic response. (B) ROC curves for different population sizes, based on a cut-off for immunogenicity of 0.4.

http://dx.doi.org/10.1371/journal.pcbi.1004796.g004

Crossreactivity.

We also investigated how the crossreactivity of the peptide affected the performance of the BIITE algorithm. Intuitively, we expect that is it easier to determine the HLA ligands of a peptide that is poorly reactive (ie, not many positive ELISpots). This is easiest to understand in the extreme case of immunogenicity values Ej∈{0,1}: each negative ELISpot would indicate that none of the HLAs of the subject is immunogenic; while a positive ELISpot result only informs us that (at least) one of the subject's HLAs is immunogenic in combination with the peptide. In this extreme case, when all the subjects produce positive ELISpots, it is impossible to attribute these positive outcomes to any HLA.

To confirm the effect of the number of positive ELISpots, we ordered the peptides in the Burkholderia dataset based on the success rate (i.e. how many positive results they produced) in the ELISpot assay in our cohort, and grouped the peptides accordingly (see S3A Fig). We can now compute the performance of the algorithm on each of these subsets. Following the reasoning above, we predict that the AUC will be negatively correlated with the success rate (S3B Fig); this is indeed the case (S3C Fig). Note that if very few positive ELISpots are observed, the algorithm might also underperform: in the extreme case of observing one positive ELISpot, there are many candidate HLAs in that subject that could be presenting the peptide. If these HLA molecules are present at similar frequencies in the cohort, the algorithm is unable to distinguish which of the possible pHLA complexes is responsible for triggering the CD4+ T cell response.

Comparison to RATE.

To our knowledge, there exists only one other algorithm, RATE [10], which uses ELISpot data to infer immunogenicity of pHLA complexes. In contrast with our Bayesian approach (which allows for inclusion of prior data and considers all HLAs simultaneously), the RATE software uses the frequentist exact Fisher test to compare the odds ratios (and the relative frequency, RF) of having a positive ELISpot in the HLA+ and HLA- groups. We first applied RATE to the Burkholderia dataset; it returned two potential pHLA restrictions with high RF: peptide5:DRB1*01 and peptide5:DQB1*04. Transgenic mouse data is only available for the first combination and was negative. Both combinations have a high (>0.9) posterior mode in our analysis, but lack power (Kullback-Leibler divergence below Beta cut-off) as both occur only twice in our cohort. The overall AUC based on RF was 53.6% (compared to 71.3% for BIITE). When applied to the Pseudomonas dataset, RATE returned 54 pHLA combinations with RF greater than 2; unfortunately, there is no transgenic mouse data available for any of these combinations. Based on the pHLA combinations that were tested in transgenic mice, we found that RATE reached an AUC of 49% (compared to 68.8% for BIITE). In the two subcohorts of the Pseudomonas dataset, RATE reached AUCs of 55% in group 1 (compared to 69% or 73% for BIITE, depending on the inclusion of a prior based on binding predictions, see above) and 29% in group 2 (compared to 32% or 69% for BIITE).

Comparison to NetMHCIIPan.

Immunogenicity and binding affinity are correlated and so it is also instructive to compare our algorithm with the best available affinity predictor, namely NetMHCIIPan [9,17]. We used NetMHCIIPan to construct ROC curves, based on the binding threshold, for the pHLA combinations we had previously considered. For the Burkholderia data, this yielded an AUC of 67.7% (compared to 71.3% from our analysis). Since we also obtained experimental binding data for several pHLA combinations from the Burkholderia dataset, we were interested in the cases where the binding assay and transgenic mouse models gave discordant results. These can be divided in two groups: one where the transgenic mouse model suggests immunogenicity, while the binding assay indicates there is no stable binding, and vice versa. In the first group there were five such observations. Only one of these (20%) is predicted to bind by NetMHCIIPan. Our algorithm is powered to determine 3 of the pHLA combinations of which two (66%) are correctly identified as immunogenic (S3A Table). The second group, which contains pHLA combinations which reach stable binding but are not immunogenic based on the transgenic mouse study, is unsurprisingly larger, with 34 such combinations. Of these, NetMHCIIPan predicts 11 (32%) to be binding (at the 500nM cut-off, see S3B Table); this represents a relatively low success rate for predicting binding but would be a high failure rate for predicting immunogenicity as none of these pHLA are immunogenic. As for our analysis, 16 combinations were informative (reached the Kullback-Leibler cut-off, S3C Table); of these, two (12.5%) were incorrectly predicted to be immunogenic. This indicates that BIITE can successfully distinguish between binding and immunogenicity. Another important benefit of BIITE compared to NetMHCIIPan is that the results of BIITE are cohort-specific. That is, immunogenicity in two cohorts (e.g. controllers and non-controllers) infected with the same pathogen can be compared using BIITE. BIITE may find the same HLA:peptide combination to be immunogenic in one cohort but not in another allowing an analysis of differing immune control between the cohorts whereas NetMHCIIPan can only be used to determine whether a particular HLA:peptide combination is ever immunogenic. This is illustrated by our work on the PA dataset. In that dataset, we could distinguish clinically between two groups which were marked by different ELISpot response rates (group 2 i.e. "non-controllers" were poorly responsive with few positive ELISpots). Since NetMHCIIPan does not use cohort data, it produces the same peptide:HLA binding predictions (which in turn serve as a proxy for immunogenicity) for both groups and hence fails to capture the real differences in immunogenicity between the two groups. If we use transgenic mouse data as the validation dataset for the poorly responsive group 2 then NetMHCIIPan reached an AUC of 81.3%, compared to BIITE's AUC of 68.8% (S4 Fig); however this analysis is misleading. The cohort produces very few positive ELISpot responses and, accordingly, BIITE identifies very few HLA:peptide combinations as being immunogenic compared to NetMHCIIPan. Arguably, BIITE's prediction of very few immunogenic peptide:HLA combinations compared to NetMHCIIPan is a more accurate representation of immunogenicity in this poorly responsive group.

Effect of HLA characteristics on model outcomes.

Two concerns might arise regarding the performance of our model: firstly, the model might prefer to assign higher immunogenicity values to HLAs which are more common, as they might 'hide' less common but immunogenic alleles in their hosts. Secondly, we want to ensure our analysis goes beyond a simple 'counting' algorithm, in which one considers immunogenic exactly those pHLA combinations for which the highest percentage of HLA carriers produce positive ELISpots. In order to confirm that the model is able to detect immunogenicity of peptide:HLA combinations where the HLA is rare in the population, we plotted the predicted immunogenicity (posterior mode) of each pHLA combination in the Burkholderia data set against the number of carriers of the HLA. While having a low number of HLA carriers is restrictive in the sense that their posterior distributions do not reach the Kullback-Leibler cut-off, those that do pass this requirement can be found to be immunogenic by the algorithm (S5A Fig). Similarly, the percentage of positive ELISpots per HLA allele is not strongly correlated with the posterior mode (S5B Fig); if we simply use the count to predict immunogenicity then the AUC is very low (30%) suggesting our algorithm is contributing far more than simply counting.

Validation on an HIV-1 dataset

As described in the methods, we also analysed the HLA-I dataset that was analysed in [10]. For each peptide, HLAs were ranked by descending posterior mode provided by BIITE. To validate the results, we compared the outcome to the LANL A-list [18] of best defined epitopes in HIV-1. This is not ideal, since a peptide:HLA combination can be immunogenic without being featured on this A-list (i.e. we will tend to overestimate false positives); nevertheless, we used this list to determine lower bounds on the AUC. Both BIITE (AUC = 81%) and RATE (AUC = 77%) performed well and were able to identify A-list peptide:HLA combinations (S6 Table). However, RATE's 'concise output file' (S7 Table) included a number of peptide:HLA combinations which are most likely not immunogenic but appear to be so due to linkage in the HLA locus (highlighted in red in S7 Table. Strong linkage: p-value chi-square test < 2.2*10E-16; linkage: p-value chi-square test < 0.01); this was not a problem with the BIITE algorithm. This shows that BIITE is successful in handling possible confounding due to linkage disequilibrium between genetic loci, a recurring problem that is not dealt with by RATE.

Discussion

Determining the immunogenicity of peptide:HLA (pHLA) combinations is an important task when studying correlates of immunity in infection, which in turn may inform vaccine design. Different experimental and computational approaches have been used to establish this. The experimental approaches include assessing the binding strength of the pHLA complex, in vitro antigen presentation studies using CD4+ T cell clones, and HLA class II- transgenic mouse models to mimic pathogen peptide expression in a human host. Each of these methods has shortcomings. While binding (whether measured in binding assays or predicted) of a peptide and immunogenicity of that peptide are correlated, binding is insufficient to guarantee immunogenicity which entails other processes both in the antigen presentation pathways (such as cleavage of protein into smaller peptides and the loading of peptides onto HLA-II molecules) but also in the selection and expansion of the T cell clone(s) bearing the relevant TCR(s) in vivo. More direct assays of immunogenicity such as CD4+ T cell cloning from bulk cultures and assessment of peptide presentation using transfectants expressing individual HLA-II heterodimers are cumbersome and intractable for large-scale studies [19,20]. Transgenic mice are good in vivo models encompassing peptide processing and presenting pathways, but are labour-intensive to develop and maintain. Furthermore, the overlap between which peptides are immunogenic in humans and in mouse models is not perfect and seems to be highly dependent on the choice of the effector cell [21]. As the mouse model incorporates a single human HLA, it is also unsuitable to appreciate the effects of immunodominance, where different HLAs may present different peptides simultaneously to the host's immune system [22].

ELISpot assays use ex vivo PBMCs from exposed hosts to test peptidic immunogenicity. In theory, this approach includes effects of peptide presentation pathway idiosyncrasies, binding affinities between peptide and HLA, and immunodominance. As a trade-off however, we lose the clarity of the experimental approaches where each HLA is studied independently, and it is difficult to assess which peptide:HLA combinations cause the observed ELISpot outcomes.

Here, we have presented a Bayesian framework to perform this analysis. We have shown its accuracy on both simulated and experimental data, and have determined the effect of sample size. While higher sample sizes (100-200 subjects) would have yielded better results, the analysis on a very limited dataset of 38 Burkholderia-sero-positive exposed subjects performed as expected (AUC = 71%) from simulated data. A caveat should be given regarding the composition of the data set: analysis on data sets with high rates of positive ELISpots suffer from poor identifiability. Nevertheless it is remarkable that with such a small dataset relatively high specificity and sensitivity could be achieved; while most algorithms for predicting immunogenicity or binding require training datasets with hundreds of data points per HLA allele, we typically have less than 10 data points per allele. Furthermore, unlike prediction algorithms, BIITE explicitly uses cohort data, which makes it suitable to interrogate differences in immunogenicity between patient groups responding differently to the same pathogen; this would be impossible for a binding prediction algorithm.

This tool is available online as the R package BIITE and is, in essence, not ELISpot or HLA class II-specific. In general, it can be used to attribute Boolean outcomes (such as therapeutic success or disease progression) to allelic variation on multiple genes. As an example, we have applied BIITE to an HLA class I ELISpot dataset (HIV-1). In general, to apply the algorithm to a new problem it is necessary to consider carefully the source of the priors if these are to be used; establish the length of the MH chain needed to obtain a representative sample of the posterior distribution; and to consider issues of sample size and collinearity in the dataset.

It also allows for the addition of prior information ad libitum: we can envisage a scenario where not only binding predictions (from NetMHCIIPan or other binding prediction algorithms) but also binding data and transgenic mouse data are included in the prior. Of course, conflicting binding information would result in different priors, and hence influence the outcome of the algorithm. However, the effect of the prior is dependent on sample size; in a suitably large dataset, the influence of the prior is outweighed by the likelihood factor. Alternatively, confidence in the prior can be expressed in the prior itself: the more 'flat' a prior, the less it will influence the posterior distribution.

In summary, we have developed a general approach which can be used for analysing complex biological data. Whether or not a peptide:HLA combination is immunogenic involves a number of interconnected immunological processes. To sidestep this complexity, reductionist experimental approaches are often taken, e.g. in vitro stimulation with APCs expressing a single HLA molecule; these approaches do not recapitulate a number of, potentially important, biological details. We have developed an approach that allows us to use the most pertinent but also most complex experimental data (ELISpot) and interrogate it to obtain accurate results.

Supporting Information

S1 Fig. Priors as used, depending on the predictions from NetMHCIIPan.

A peptide:HLA combination is considered to be a binder if the predicted binding affinity is below 500nM; the prior for such a combination is the Beta distribution with mode 0.35 and SD 0.2. Peptide:HLA combination with predicted binding affinity above 500nM have the Beta distribution with mode 0.001 and SD 0.15 as prior.

(PNG)

S2 Fig. A comparison of the odds ratios of having a positive ELISpot (i.e. number of positive ELISpots/number of negative ELISpots for that peptide) in the original Burkholderia dataset and the simulated data.

(PNG)

S3 Fig. Datasets with high ELISpot success rates are harder to analyse.

(A) We ordered the Burkholderia peptides based on their ELISpot success rates and considered subsets of this dataset. We predict the AUCs to be negatively correlated with the success rate, i.e. the subsets to be ordered as shown in (B) (subset 8 predicted to have the lowest AUC, subset 1 the highest). This is indeed what we observe in (C).

(PNG)

S4 Fig. Comparison to NetMHCIIPan.

(A) For the Burkholderia data. (B) for the Pseudomonas data.

(PNG)

S5 Fig. The algorithm avoids the following pitfalls.

(A) Posterior modes are not correlated with the number of HLA carriers. (B) Posterior modes are not strongly correlated with the percentage of HLA carriers that produced a positive ELISpot. Data is taken from the analysis of the Burkholderia data set.

(PNG)

S1 Table. Count and frequency of DRB1 and DQB1 alleles in the Burkholderia cohort.

(DOCX)

S2 Table. Count and frequency of DRB1 and DQB1 alleles in the Pseudomonas cohort.

(DOCX)

S3 Table. Discordant results between binding assay and transgenic mouse models in Burkholderia data set.

(A) Peptide:MHC combinations which were found to be immunogenic in the transgenic mouse model but did not bind in the binding assay, together with their predicted binding IC50 from NetMHCIIPan, and the analysis of the posterior mode from our analysis. The Kullback-Leibler divergence of the marginal posterior with respect to the uniform distribution is also indicated. (B,C) Peptide:MHC combinations which were not immunogenic in the transgenic mouse assays but showed stable binding in the binding assay. Of these 34 combinations, 11 were predicted to be binding (IC50 < 500 nM) by NetMHCIIPan (B). Marginal posterior distributions from our model passed the Kullback-Leibler cut-off for 16/34 combinations (C). Between these two tables, only five were found in common (highlighted in matching colours).

(DOCX)

S4 Table. Example of an input file.

(CSV)

S5 Table. Count and frequency of HLA-A, -B and -C alleles in the HIV-1 cohort.

(DOCX)

S6 Table. Analysis of the HIV-1 ELISpot dataset (BIITE).

We analysed data from 32 peptides (first column). These 32 peptides were the ones studied in [10] and were chosen because they contain LANL HIV-1 A list epitopes that are restricted by HLAs which are abundant in the cohort; this HLA is in the column 'Restricting HLA (RATE)'. We also include the p value reported by RATE for each peptide:HLA combination. In the next column, 'Restricting HLA (BIITE)', we list the HLAs identified using BIITE that appear to elicit CD8+ T cell responses in combination with the peptide. For these HLAs, we provide the number of carriers that presented a positive ELISpot response for the peptide in question and we count how in how many responding carriers of the HLA this HLA had the highest posterior mode ('# Carriers with positive ELISpot and highest rank'). We also show the posterior mode attributed to the peptide:HLA combination. The last two columns document whether (and where) the peptide:HLA combination was found in the LANL database.

(XLSX)

S7 Table. Analysis of the HIV-1 ELISpot dataset (RATE).

Peptide:HLA combinations in the 'concise results' of RATE. Green highlighted alleles were detected by BIITE, while red highlighted alleles are probably not immunogenic in combination with the peptide, but appear to be so because of linkage disequilibrium.

(XLSX)

S8 Table. Example of an output file (pep_1 from the simulated data).

HLAs are ranked from highest to lowest posterior mode ('Posterior Mode') and we provide the number of carriers in the cohort ('# Carriers') and the number of carriers who responded to the peptide. We also show, per HLA, how many carriers have this HLA as their highest ranked HLA ('# Carriers with highest rank'); and how many carriers who responded to the peptide have this HLA as their highest ranked HLA ('# Carriers with positive ELISpot and highest rank'). This last column can be interpreted as the number of positive ELISpots that are to be attributed to the specific HLA.

(XLSX)

S1 Text. Mathematical considerations.

An explanation of the rationale and machinery behind BIITE aimed at readers who do not have a background in Bayesian analysis.

(DOCX)

Acknowledgments

We have used the high performance cluster at Imperial College for this project. The authors thank the Respiratory Biomedical Research Unit and patients at Royal Brompton and Harefield National Health Service Foundation Trust funded by National Institute for Health Research. We thank Professor Robert Wilson, Professor Diana Bilton and Dr. Michael Loebinger at Royal Brompton Hospital London for their help recruiting the bronchiectasis/PA clinical cohorts. We thank the staff members and patients at the Blood Transfusion Centre, Khon Kaen Hospital, Khon Kaen, Thailand.

Author Contributions

Conceived and designed the experiments: LB DMA RJB BA PKO. Performed the experiments: LB. Analyzed the data: LB BA. Contributed reagents/materials/analysis tools: KJQ CJR BM JHR GL. Wrote the paper: LB BA DMA RJB.

References

Yewdell JW. Confronting complexity: real-world immunodominance in antiviral CD8+ T cell responses. Immunity. 2006;25: 533-43. doi: 10.1016/j.immuni.2006.09.005. pmid:17046682

Czerkinsky CC, Nilsson L-Å, Nygren H, Ouchterlony Ö, Tarkowski A. A solid-phase enzyme-linked immunospot (ELISPOT) assay for enumeration of specific antibody-secreting cells. J Immunol Methods. 1983;65: 109-121. doi: 10.1016/0022-1759(83)90308-3. pmid:6361139

Charron DJ, Lotteau V, Turmel P. Hybrid HLA-DC antigens provide molecular evidence for gene trans-complementation. Nature. 1984;312: 157-159. doi: 10.1038/312157a0. pmid:6209579

Gjertsen HA, Lundin KE, Ronningen KS, Gaudernack G, Thorsby E. T cells recognizing an HLA-DQ alpha beta heterodimer encoded in cis by the DR4DQw4 haplotype and in trans by DR4DQw8/DRw8DQw4 heterozygous cells. Hum Immunol. 1991;30: 226-232. pmid:2055788 doi: 10.1016/0198-8859(91)90039-c

Kwok WW, Schwarz D, Nepom BS, Hock RA, Thurtle PS, Nepom GT. HLA-DQ molecules form alpha-beta heterodimers of mixed allotype. J Immunol. 1988;141: 3123-3127. pmid:2844907

Altmann DM. HLA-DQ associations with autoimmune disease. Autoimmunity. 1992;14: 79-83. pmid:1299350 doi: 10.3109/08916939309077360

Lotteau V, Teyton L, Burroughs D, Charron D. A novel HLA class II molecule (DR alpha-DQ beta) created by mismatched isotype pairing. Nature. 1987;329: 339-41. doi: 10.1038/329339a0. pmid:3306409

Lotteau V, Sands J, Teyton L, Turmel P, Charron D, Strominger JL. Modulation of HLA class II antigen expression by transfection of sense and antisense DR alpha cDNA. J Exp Med. 1989;169: 351-356. pmid:2462612 doi: 10.1084/jem.169.1.351

Karosiene E, Rasmussen M, Blicher T, Lund O, Buus S, Nielsen M. NetMHCIIpan-3.0, a common pan-specific MHC class II prediction method including all three human MHC class II isotypes, HLA-DR, HLA-DP and HLA-DQ. Immunogenetics. 2013;65: 711-24. doi: 10.1007/s00251-013-0720-y. pmid:23900783

Paul S, Dillon MBC, Lindestam Arlehamn CS, Huang H, Davis MM, McKinney DM, et al. A population response analysis approach to assign class II HLA-epitope restrictions. J Immunol. American Association of Immunologists; 2015;194: 6164-76. doi: 10.4049/jimmunol.1403074

Hastings WK. Monte Carlo sampling methods using Markov chains and their applications. Biometrika. 1970;57: 97. doi: 10.2307/2334940

R Core Team. R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing; 2012

Wickham H. ggplot2. Springer. 2009

Reynolds C, Goudet A, Jenjaroen K, Sumonwiriya M, Rinchai D, Musson J, et al. T Cell Immunity to the Alkyl Hydroperoxide Reductase of Burkholderia pseudomallei: A Correlate of Disease Outcome in Acute Melioidosis. J Immunol. American Association of Immunologists; 2015;194: 4814-24. doi: 10.4049/jimmunol.1402862

Quigley KJ, Reynolds CJ, Goudet A, Raynsford EJ, Sergeant R, Quigley A, et al. Chronic Infection by Mucoid Pseudomonas aeruginosa Associated with Dysregulation in T-Cell Immunity to Outer Membrane Porin F. Am J Respir Crit Care Med. American Thoracic Society; 2015;191: 1250-64. doi: 10.1164/rccm.201411-1995OC

Kiepiela P, Ngumbela K, Thobakgale C, Ramduth D, Honeyborne I, Moodley E, et al. CD8+ T-cell responses to different HIV proteins have discordant associations with viral load. Nat Med. Nature Publishing Group; 2007;13: 46-53. doi: 10.1038/nm1520

Nielsen M, Lund O, Buus S, Lundegaard C. MHC class II epitope predictive algorithms. Immunology. 2010;130: 319-28. doi: 10.1111/j.1365-2567.2010.03268.x. pmid:20408898

Llano A, Williams A, Olvera A, Silva-Arrieta S, Brander C. Best-Characterized HIV-1 CTL Epitopes: The 2013 Update. In: Yusim K, Korber BTM, Brander C, Barouch D, Boer de R, Haynes BF, et al., editors. HIV Molecular Immunology 2013. Los Alamos National Laboratory, Theoretical Biology and Biophysics; 2013. pp. 3-25

Wilkinson D, de Vries RR, Madrigal JA, Lock CB, Morgenstern JP, Trowsdale J, et al. Analysis of HLA-DR glycoproteins by DNA-mediated gene transfer. Definition of DR2 beta gene products and antigen presentation to T cell clones from leprosy patients. J Exp Med. 1988;167: 1442-1458. pmid:3128633 doi: 10.1084/jem.167.4.1442

Pette M, Fujita K, Wilkinson D, Altmann DM, Trowsdale J, Giegerich G, et al. Myelin autoreactivity in multiple sclerosis: recognition of myelin basic protein in the context of HLA-DR2 products by T lymphocytes of multiple-sclerosis patients and healthy donors. Proc Natl Acad Sci U S A. 1990;87: 7968-7972. pmid:1700423 doi: 10.1073/pnas.87.20.7968

Kotturi MF, Assarsson E, Peters B, Grey H, Oseroff C, Pasquetto V, et al. Of mice and humans: how good are HLA transgenic mice as a model of human immune responses? Immunome Res. 2009;5: 3. doi: 10.1186/1745-7580-5-3. pmid:19534819

Deng H, Apple R, Clare-Salzler M, Trembleau S, Mathis D, Adorini L, et al. Determinant capture as a possible mechanism of protection afforded by major histocompatibility complex class II molecules in autoimmune disease. J Exp Med. 1993;178: 1675-1680. pmid:8228814 doi: 10.1084/jem.178.5.1675