Background
Identification of HLA alleles restricting specific T cell epitopes is an important component of accurate characterization of T cell responses. This information is required, for example, for the production of tetramer staining reagents [
1‐
3], or to evaluate association of particular HLAs with protective or predisposing T cell responses [
4‐
6]. The restricting HLA alleles can be determined by experiments relying on classical immunological approaches, such as inhibition by HLA locus specific antibodies, and use of matched/mismatched or single HLA allele transfected cell lines [
7]. These experimental approaches can be time consuming and resource intensive. As an alternative, we developed a computational method called RATE (Restrictor Analysis Tool for Epitopes) that infers HLA restriction of epitopes from T cell response data in HLA typed subjects [
8]. “T cell response data” is the specific immune response to various epitopes in PBMCs from HLA typed individuals measured by, for example, IFN-γ ELISPOT and reported as spot-forming cells (SFCs) per million cells. RATE infers HLA restrictions by considering the presence or absence of a response to a given epitope as the biological outcome, and calculating the relative frequency (RF) of the subjects responding to a given epitope and expressing a given allele as compared to the general test population and associated statistical significance.
This method was initially validated with a small set of experimental data, generated to verify a limited number of inferred restrictions, and by retrospective analysis of data sets publicly available online. We recently reported the results of a clinical study for which HLA restrictions were experimentally determined for 191
Mycobacterium tuberculosis (MTB) peptides tested in a South African cohort of 63 MTB infected individuals by the use of single HLA transfectants [
9]. This provided an opportunity for an unbiased validation of RATE, and also for further optimization of its performance by systematically examining the effect of varying different parameters linked to the analysis and its output. The subsequently updated version of the RATE tool server has also been made available online (
http://iedb-rate.liai.org/).
Materials and methods
Study subjects and peptides
The study involved MTB-specific T cell response data from healthy adults with latent MTB infection from the Worcester region of the Western Cape Province of South Africa, as detailed in Arlehamn et al. (2016) [
9]. The responses studied are resulting from natural exposure to whole TB. MTB donors were recruited based on IGRA (Interferon gamma release assay; FDA approved for diagnosis of latent TB infection) reactivity and lack of active TB symptoms. Donors with a positive IGRA are latently infected with MTB. Peptides representing the vaccine candidate and IGRA antigens (Rv3874; CFP10 and Rv3875; ESAT-6) (15-mers overlapping by 10 amino acids spanning each entire protein) and epitopes from the frequently recognized antigens previously reported by Arlehamn et al. [
10], as well as additional frequently recognized epitopes described in ex vivo experiments and available in the IEDB database (
www.iedb.org) [
11‐
15] were included in the study as described previously [
9].
Data on allergen-epitope T cell reactivity independently and previously reported from a separate cohort of Timothy grass (TG) allergic donors from the Denver, CO and San Diego, CA regions was also investigated. Donors had a skin prick test of > 3 mm to Timothy grass or a TG-specific IgE titer of > 0.35 kU/L and a clinical history of seasonal allergic symptoms consistent with Timothy grass pollen allergy [
16‐
19]. Immunodominant peptides from Timothy grass pollen T cell antigens were studied, as well as peptides from other grasses, including Kentucky blue grass, Rye grass, Canary grass and Orchard grass. These peptides were conserved in grass pollen across species and elicited responses in two or more pollen allergic patients [
16‐
19]. Peptides were synthesized as crude material on a small (1 mg) scale by A and A (San Diego).
PBMC isolation, ELISPOT assays and HLA typing
Peripheral blood mononuclear cells (PBMC) were purified from whole blood by layering onto Ficoll and density-gradient centrifugation, according to the manufacturer’s instructions.
Cells were cryopreserved in liquid nitrogen suspended in FBS containing 10% (vol/vol) DMSO.
For ELISPOT assays, PBMC were stimulated at 2 × 105 cells/well in triplicate with peptide pools (5 μg/ml), peptides (10 μg/ml), PHA (10 μg/ml) or medium containing 0.25% DMSO (percent DMSO in the pools, as a control) in 96-well plates (Immobilion-P; Millipore) coated with 5 μg/ml anti-IFNγ (1-D1K; Mabtech). After 20 h incubation at 37 °C, wells were washed with PBS/0.05% Tween 20 and incubated with biotinylated anti-IFNγ (7-B6-1; Mabtech) for 2 h. Spots were developed using Vectastain APC peroxidase (Vector Laboratories) and 3-amino-9-ethylcarbazole (Sigma-Aldrich). Spots were counted by computer-assisted image analysis (KS-ELISPOT reader; Zeiss). Responses were considered positive if the net spot-forming cells (SFC) per 106 PBMC were ≥20, the stimulation index ≥2, and p ≤0.05 (Student’s t-test, mean of triplicate values of the response against relevant pools or peptides vs. the DMSO control). All samples had a viability >75%, as determined by trypan blue, and reactivity to PHA >400 SFC/106 cells.
Four-digit HLA typing for these cohorts was done as previously described [
20]. Genomic DNA was isolated from PBMC using standard techniques (REPLI-g; Qiagen). Amplicons for HLA class I and class II genes were generated using PCR and locus-specific primers. Amplicons of the correct size were purified using Zymo DNA Clean-up Kit, according to the manufacturer’s instructions. Sequencing libraries were prepared using Nextera XT reagents (Illumina), according to manufacturer’s instructions. The libraries were purified using AMPure XP (Beckman Coulter) with a ratio of 0.5:1 beads to DNA (vol/vol). The libraries were pooled in equimolar amounts and loaded at 5.4pM on one MiSeq flowcell with 1% phiX spiked in (MiSeq Reagent Kit v3). Paired-end sequencing was performed with 300 cycles in each direction. HLA typing calls were made using HLATyphon (
https://github.com/LJI-Bioinformatics/HLATyphon).
HLA restriction using single HLA transfected cell lines
HLA restriction assays using single HLA transfected cell lines were performed as described earlier [
9]. Single HLA transfected RM3 (derived from human B lymphocyte cell line Raji) or DAP.3 (L cell fibroblast) were maintained in culture. In preparation for the assay, the cell lines were harvested and viability (all >75%) was determined using Trypan Blue. Each cell line at 2x10
5 cells/well was pulsed with 10 μg/ml individual peptide for 1 h at 37 °C, followed by four washes in RPMI. PBMC at 2x10
5/well were stimulated in triplicate with peptide pulsed cell line (5x10
4 cells/well), cell line alone (as a control), peptides (10 μg/ml), PHA (10 μg/ml) or medium containing 0.25% DMSO (percent DMSO in the peptides, as a control) in 96-well plates (Immobilion-P; Millipore) coated with anti-IFNγ antibody as described above for single cytokine ELISPOT. Criteria for positive responses were as described for ELISPOT assays above.
RATE calculations
The RATE tool (
http://iedb-rate.liai.org/) [
8] was used to computationally infer the HLA restrictions from the immune response and HLA typing data described above. RATE estimates Relative Frequency (RF) to quantify the strength of associations between expression of a specific allele and detection of positive immune response. An RF > 1 indicates a positive association between the two properties in question (i.e., expressing the specific allele increases the “odds” of having positive immune response). RF is calculated according to the formula:
$$ R F=\frac{A^{+}{R}^{+}/\left({A}^{+}{R}^{+}+{A}^{+}{R}^{-}\right)}{\left({A}^{+}{R}^{+} + {A}^{-}{R}^{+}\right)/ Total\ donors} $$
Where
A
+
R
+
= Number of subjects who expressed a specific allele and gave a positive immune response to the specific peptide
A
-
R
-
= Number of subjects who did not express the specific allele and did not give a positive immune response to the specific peptide
A
-
R
+
= Number of subjects who did not express the specific allele but gave a positive immune response to the specific peptide
A
+
R
-
= Number of subjects who expressed the specific allele but did not give a positive immune response to the specific peptide
The Fisher’s exact test is used to estimate the statistical significance of the association between HLA molecules and epitope responses.
Statistical evaluation of RATE results
In order to evaluate the performance of the RATE tool, the following statistical measures were estimated:
a)
Matthew’s Correlation Coefficient
$$ M C C=\frac{\left( TP\times TN\right)-\left( FP\times FN\right)}{\sqrt{\left( TP+ FP\right)\left( TP+ FN\right)\left( TN+ FP\right)\left( TN+ FN\right)}} $$
Where
TP = True positives
FP = False positives
FN = False negatives
TN = True negatives
b)
Accuracy
$$ Accuracy=\frac{TP + TN}{Total} $$
c)
Sensitivity
$$ Sensitivity=\frac{TP}{TP+ FN} $$
d)
Specificity
$$ Specificity=\frac{TN}{FP+ TN} $$
e)
Precision
$$ Precision=\frac{TP}{TP+ FP} $$
f)
False positive rate
$$ F P R=\frac{FP}{FP+ TN} $$
Discussion
Here, we have utilized experimental data generated from an independent epitope identification study to validate and further optimize the performance of the RATE tool [
8], developed earlier to infer HLA restrictions based on HLA typing and immune response data in human populations. Specifically, the present study takes advantage of a recently described data set, where HLA restrictions were experimentally determined for a set of 191 different MTB derived peptides tested in 63 MTB infected South African donors. We found that, on this data set, RATE was associated with a performance of MCC = 0.451 when optimal cutoff values for the output parameters were applied. Furthermore, the tool was associated with an accuracy of 0.745 and sensitivity of 0.550. In a practical sense, this performance indicates that the tool would allow a user to greatly reduce the number of potential restrictions to be examined, while still identifying about half of the true restrictions without any experimental work. The reason for the relatively low sensitivity is likely due to the fact that several restrictions occur infrequently and are thus not detected by an association based approach. In this respect, the fundamental utility of RATE from the viewpoint of an experimental user is that it identifies the most frequently occurring restrictions, thereby facilitating more efficient use of precious laboratory reagents and donor samples for subsequent analyses.
In utilizing the experimental data set to optimize tool performance, first consideration was given to the reliability of determinations as judged by the associated p-value in a Fisher’s exact test. Perhaps not surprisingly, we found optimal tool performance by considering only restrictions associated with a p-value <0.01. Interestingly, the performance of the tool is decreased by imposing significance levels less than 0.01. This has particular significance in terms of the potential use of Bonferroni correction, which we considered in the RATE tool output. The results clearly indicated that a Bonferroni correction should not be used, as it would not improve the tool performance, but rather essentially result in no useful inferences.
Imposing an additional requirement for an A+R+ threshold which result in selection of more reliable inferences, namely those HLA/epitope combinations based on epitopes recognized by multiple donors expressing a specific allele improved the performance compared to p-value cutoff alone. For this particular data set the MCC improved to 0.497 when A+R+ ≥ 5 cutoff was applied along with p <0.05. However, we do not recommend using this as a general threshold, because the optimal A+R+ threshold is expected to be strongly dependent on the absolute number of donors associated with a particular data-set. This type of filter could nevertheless be considered and adjusted to fit the experimental context, such as when a relatively large number of inferred restrictions can be feasibly tested, or whether it is desired to test only few higher probability candidates.
Following a different approach, we saw that increasing the magnitude of the associations, as measured by RF values improve RATE performance. In the present analysis we empirically determined and applied an optimal performance for RF value of 1.3. The RF threshold can be easily adjusted if more or less stringency is desired. We further emphasize how adjusting the RF value is indirectly correlated with p and A+R+ values. For this reason, in most cases adjusting the p-value threshold will also implicitly select for higher RF or A+R+ values as well.
In terms of further refinements, surprisingly we found that incorporating the predicted HLA binding in the restriction scheme was optimal when used in isolation for an IEDB consensus percentile rank of 15.0 but did not improve the performance in a broad range of percentile ranks (5.0 to 25.0) when used in combination with other optimized parameters. Several different factors might contribute to this result. First, it is well established that HLA binding is a necessary but not sufficient requirement for T cell recognition; in the case of HLA restrictions, all peptides studied are by definition binders to some of the alleles, and the well-known promiscuity observed in the case of HLA class II binding [
22] might hinder realizing any increase in RATE performance based on HLA binding predictions. Second, it is possible that the result reflects that HLA class II binding predictions for certain alleles may be relatively inaccurate. This concern will be addressed in future by progressive retraining of HLA class II prediction tools, as more HLA binding data becomes available, and increased accuracy can be achieved.
We also found that the iterative combination of different allele subsets described in the previous study [
8] did not improve RATE performance. However, in the MTB data set that was used to optimize the performance only 6 out of the total 191 peptides had promiscuous restrictions and for this reason this data set was not ideal to address the best strategy to identify promiscuous restrictions. Future studies utilizing a larger number of experimentally determined promiscuous restrictions will be required to fully evaluate this issue. At the same time it should be considered that loss of significance due to multiple comparisons is a serious problem for the promiscuous option. To truly demonstrate promiscuous restrictions might require larger data sets than the one utilized here (which is representative of most epitope identification studies). Based upon these considerations it is recommended that the “monogamous restriction” calculation be used for practical purposes (monogamous refers to HLA-peptide relationship where a peptide is found to be restricted by a single allele). To demonstrate or identify potential promiscuous restrictions it seems safer to record different HLA restrictions independently identified for a given epitope.
Finally the RATE tool was applied to infer additional restrictions both in the original data set, and in a data set including epitopes derived from pollen allergens. In the present study, we adjusted the RATE parameters according to a known MTB data set. In future studies it will be interesting to assess the performance of RATE on unknown samples to exclude overfitting of the parameters to the MTB data. However, these experiments are laborious and expensive and therefore beyond the scope of the current study.
The results highlight how the RATE approach is suited for inference of restrictions for which no transfected cell lines are available. We emphasize that these instances most often correspond to alleles that are rare in the general population, but relatively frequent in a specific study population, ethnicity or geographical location. In this respect, it is notable that several of the new restrictions inferred by the RATE tool in the MTB data set were mediated by the HLA DRB1*15:03 allele, which is present at 0.0517 frequency in the Western Cape region study population, 0.0596 frequency in South Africa, but only at 0.0225 worldwide (
http://www.allelefrequencies.net [
23]). These results emphasize the value of the RATE tool in terms of providing HLA restriction data in the context of diverse HLAs and complex multi-ethnic human trials.