Estimating HLA-DRB1 promiscuity level

Given that large-scale experimental assays to measure individual peptide–MHC interactions are extremely tedious, we first employed established bioinformatics tools to predict the binding affinities of experimentally verified epitope peptides for a panel of 162 nonsynonymous HLA-DRB1 alleles, all of which are present at detectable frequencies in at least one human population [35–37]. The set of investigated epitopes was derived from the Immune Epitope Database (IEDB) and contains 2,691 peptide epitopes of 71 pathogens known to be bound by certain HLA class II alleles [38] (S1 Data). Epitopes showing high levels of amino acid similarity to each other were excluded from the analysis (See Methods). Most included epitopes are 15 to 20 amino acids long and are found in only one of the 71 pathogens (S1 Fig, S1 Data). The NetMHCIIpan algorithm was used to predict individual epitope–MHC interactions [30], not least because it outperforms other prediction algorithms [31]. The breadth of epitope-binding repertoire or, shortly, the level of promiscuity of individual HLA-DRB1 alleles was estimated as the fraction of epitopes with a binding affinity stronger than 50 nM to the given MHC molecule. This threshold corresponds to high-affinity binding, which is frequently observed in MHC molecules displaying immunodominance [39]. We found large variation in promiscuity levels across HLA-DRB1 alleles (S2 Fig, S2 Data). Using a smaller dataset with information from both approaches, we show that the computationally predicted and the empirically estimated promiscuity values are strongly correlated with each other (Spearman’s rho: 0.78, P = 0.004, S3 Fig). Moreover, our results are robust to changes in the affinity threshold (S4A to S4C Fig), usage of other prediction algorithms (S4D Fig), and variation in the epitope data employed (S4E Fig). As expected, promiscuous HLA-DRB1 alleles can present epitopes from a broader range of pathogen species (S5 Fig). Reassuringly, there was no correlation between allele promiscuity values and the amount of data per allele used for the training of the algorithm (Spearman’s rho: −0.38, P = 0.21).

Global distribution of promiscuous HLA-DRB1 alleles

Taking advantage of the confirmed reliability of computational predictions, we next investigated the geographic distribution of HLA-DRB1 alleles. We first collected high-quality HLA-DRB1 allele prevalence data of 96 human populations residing in 43 countries from two databases and an article [35–37]. The weighted average of promiscuity level in each population was calculated based on the promiscuity values and allele frequencies of individual alleles in the population (See Methods). The analysis revealed a large variation in mean promiscuity across geographical regions and the corresponding human populations (S1 Table). Importantly, several distantly related but highly promiscuous alleles contribute to this pattern (S1 Table). Notably, an especially high allelic promiscuity level was found in Southeast Asia, an important hot spot of emerging infectious diseases [40]. To minimize any potential confounding effect of high genetic relatedness between neighboring populations, we merged populations with similar HLA allele compositions for all further analyses (See Methods).

Link between pathogen diversity and HLA-DRB1 promiscuity level

Using the Global Infectious Diseases and Epidemiology Network (GIDEON), we compiled a dataset on pathogen richness in the corresponding 43 geographic regions [41]. It consists of 95 diseases caused by 168 extracellular pathogens, including diverse bacterial species, fungi, protozoa, and helminthes. Using the same protocol, we additionally compiled a dataset on the prevalence of 149 diseases in the same regions caused by 214 viral and other obligate intracellular pathogens. The dataset and methodology employed for the analysis are standardized and have been used previously in similar contexts [7, 8, 42].

We report a strong positive correlation between extracellular pathogen diversity and mean promiscuity: HLA-DRB1 alleles that can bind epitopes from a broader range of pathogens are more likely to be found in regions of elevated pathogen diversity (Fig 1A). This pattern is unlikely to be explained by confounding factors, such as country size or HLA-DRB1 genetic diversity across countries (S2 Table). By contrast, we found no significant association between HLA-DRB1 promiscuity level and diversity of intracellular pathogens (Fig 1B). We conclude that the geographical distribution of promiscuous HLA-DRB1 alleles has been mainly shaped by the diversity of extracellular pathogens.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Relationship between epitope-binding promiscuity and pathogen diversity.

Normalized population-level promiscuity of HLA-DRB1 alleles is shown as the function of extracellular pathogen diversity, as approximated by species count. Promiscuity scores were calculated based on standardized (i.e., z-score) allele promiscuity values and were averaged in each population group (see Methods). Significant correlations were found between extracellular pathogen species count and (A) predicted allele promiscuity level in 37 groups (Spearman’s rho: 0.5, P = 0.002) and (C) in vitro promiscuity level in 28 groups (Spearman’s rho: 0.7, P = 3*10⁻⁵). No significant correlation was found between intracellular pathogen diversity and (B) predicted and (D) in vitro promiscuity at the HLA-DRB1 locus (Spearman’s rho: 0.04 and 0.21, P = 0.81 and 0.29, respectively). Dashed lines indicate smooth curve fitted using cubic smoothing spline method in R (see Methods). Population groups were created using the 15th percentile genetic distance cutoff (see Methods). For results obtained upon using alternative distance cutoff values, see S3 Data. The underlying data for this figure can be found in S4 Data.

https://doi.org/10.1371/journal.pbio.3000131.g001

The above results hold—and are even stronger—when estimates on promiscuity were derived from empirical in vitro MHC binding data (shortly, in vitro promiscuity), downloaded from the IEDB database [38] (Fig 1C and 1D, S2 Table and S3 Data). However, these results do not exclude the possibility that the geographical link between pathogen diversity and promiscuity is indirect. More direct support on the causal relationship between the two variables comes from analysis of prior human genetic studies. To investigate this issue, we focused on two geographically widespread allelic groups with exceptionally high (HLA-DRB1*12) and exceptionally low (HLA-DRB1*03) promiscuity values, respectively, and conducted literature mining on their reported associations with infectious diseases (S3 Table). As expected, HLA-DRB1*12 was associated with protection against at least five infectious diseases, while HLA-DRB1*03 was associated with susceptibilities to eight infectious diseases, which is highly unlikely by chance (Fisher test, P = 0.003) (S3 Table, S5 Data). The data also indicate local adaptation towards elevated promiscuity under diverse pathogen pressure. The HLA-DRB1*12:02 allele is prevalent in specific regions of Southeast Asia. Compared with other alleles detected in this region, HLA-DRB1*12:02 has a relatively high promiscuity value (top 20%, Fig 2A, S2 Data). The high frequency of HLA-DRB1*12:02 has been previously suggested to reflect pathogen-driven selection during the migration of a Mongolian population to South China [43]. Indeed, this allele is associated with protection from recurrent pulmonary tuberculosis, recurrent typhoid fever, and hepatosplenic schistomiasis (S5 Data), all of which are endemic diseases in Southeast Asia [44–46]. Remarkably, the frequency of this allele increases with extracellular pathogen diversity in this region (Fig 2B). Together, these observations support the hypothesis that promiscuous epitope binding of HLA-DRB1 alleles is favored by selection when extracellular pathogen diversity is high.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. HLA-DRB1*12:02 allele promiscuity level and extracellular pathogen diversity in Southeast Asia.

(A) HLA-DRB1*12:02 has an exceptionally high promiscuity level compared with other alleles. The figure shows alleles with at least 10% frequency in at least one population in Southeast Asia. Predicted allele promiscuity values are shown. (B) The mean frequency of DRB1*12:02 increases with extracellular pathogen diversity across population groups (Spearman’s rho: 0.57, P = 0.017). Populations that resided in China, Japan, South Korea, Indonesia, Malaysia, and the Philippines were included in the analysis. Red curve indicates smooth curve fitted using cubic smoothing spline method in R (see Methods). The underlying data for this figure can be found in S4 Data.

https://doi.org/10.1371/journal.pbio.3000131.g002

Evolution of promiscuous HLA alleles

An important unresolved issue is how promiscuity has changed during the course of human evolution. Under the assumption that local pathogen diversity drives the evolution of epitope recognition of HLA class II alleles, promiscuity as a molecular trait should have evolved rapidly as human populations expanded into new territories. To investigate this issue, we combined an established phylogeny of HLA-DRB1 alleles [47] with predicted epitope-binding promiscuity values. We found that alleles with a high promiscuity level have a patchy distribution across the tree (S6 Fig), indicating that high promiscuity has multiple independent origins. To investigate this observation further, we selected a set of 96 HLA-DRB1 alleles with a detectable frequency in at least one human population and appropriate sequence data (see Methods). A comparison of all pairs of these alleles revealed that even very closely related alleles show major differences in promiscuity levels (Fig 3A). For example, alleles belonging to the HLA-DRB1*13 group show over 98% amino acid sequence identity to each other, but display as much as 57-fold variation in the predicted promiscuity levels. We conclude that the switch between high and low promiscuity levels has occurred repeatedly and in a rapid manner during the allelic diversification of the HLA-DRB1 locus.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Promiscuity changes rapidly during evolution and might be a selectable trait.

(A) For all pairs of selected alleles, the predicted promiscuity difference between two HLA-DRB1 alleles is shown as a function of amino acid distance measured after excluding the epitope-binding region. Large differences in promiscuity can be observed even between closely related pairs of alleles (e.g., at zero amino acid distance). As a result, there is no correlation between amino acid distance and promiscuity fold difference (Spearman’s rho = 0.02, P = 0.19). Amino acid distances were binned as shown on the figure (n = 308, 1,168, 564, 654, 1,492). Violin plots show the density function of promiscuity fold difference values for allele pairs in the given bin. White circles show median values; bold black lines show the interquartile range. (B) Sequence variability of an amino acid site in the epitope-binding region of HLA-DRB1 (measured as Shannon entropy) correlates positively with the site’s promiscuity fragility, measured as the median predicted promiscuity fold difference caused by a random amino acid change at the given site (see inset, Spearman’s rho: 0.76, P = 0.0001). Sites that have a larger impact on promiscuity are more diverse in human populations. Line in inset represents linear regression between the two variables. The same result was obtained when promiscuity fragility was calculated based on nucleotide substitutions instead of amino acid substitutions (Spearman’s rho: 0.73, P = 0.0004, see Methods) or when sequence variability was measured as nonsynonymous nucleotide diversity (π_A) instead of sequence entropy (S7 Fig). Sites under positive selection as identified by Furlong and colleagues [48] show significantly higher promiscuity fragility (Wilcoxon rank sum test, P = 0.0012) and are marked with asterisks (see also S8 Fig). The underlying data for this figure can be found in S4 Data.

https://doi.org/10.1371/journal.pbio.3000131.g003

We next asked how selection on promiscuity has shaped the genetic diversity along the epitope-binding region of HLA molecules. To quantify protein sequence variability at each amino acid position, we calculated the Shannon entropy index based on the alignment of the 96 selected HLA-DRB1 alleles from above. For each position, we also calculated promiscuity fragility, that is, the median impact of single amino acid substitutions on promiscuity (see Methods). A strong positive correlation was found between Shannon entropy and promiscuity fragility (Fig 3B, Spearman’s rho = 0.76, P = 0.0001). Importantly, this conclusion does not depend on how sequence polymorphism was estimated (S7 Fig). Accordingly, amino acid positions with a large impact on epitope-binding promiscuity are highly variable in human populations. Furthermore, those sites in the epitope-binding region that are under positive selection [48] tend to have high promiscuity fragility values (Wilcoxon rank sum test, P = 0.0012, Fig 3B; see also S8 Fig). The above data suggest a link between allele promiscuity and HLA-DRB1 diversification, probably as an outcome of selection for locally optimized promiscuity levels.

Finally, we note that several variable molecular sites in the binding region of HLA-DRB1 affect epitope-binding characteristics without any major impact on promiscuity per se. For example, our computational analysis indicates that mutations at amino acid site numbers 9 and 47 do not seriously affect promiscuity level (Fig 3B). However, several mutations at these sites are associated with binding self-peptides and thereby shape vulnerability to specific autoimmune diseases [49, 50].

Pathogen diversity and HLA class I promiscuity

The relationship between pathogen diversity and epitope-binding promiscuity may be more general, as similar results hold for the HLA-A locus. HLA-A is one of the three types of classical human MHC class I molecules and is mainly involved in the presentation of epitopes from intracellular pathogens [51]. In agreement with expectation, we report a positive correlation between local intracellular pathogen diversity and the HLA-A promiscuity level of the corresponding human populations (S9A and S9B Fig, S3 Data). No marked positive correlation was found for two other MHC class I genes (HLA-B and HLA-C, see S9C to S9F Fig, and S3 Data). Therefore, other unrelated evolutionary forces may shape the geographical distribution of promiscuous HLA-B and HLA-C alleles (S1 Text).

Computational prediction of epitope-binding promiscuity

The IEDB has collected the results of individual and systematic studies on epitope binding by MHC alleles [38]. The experimental studies include HLA-binding assays, T-cell activation assays, and immunopeptidomic studies as well. Epitopes of all available viral, bacterial, and eukaryotic pathogens known to be bound by at least one HLA-I or HLA-II allele were extracted from IEDB. Reference proteomes of pathogenic species that carry at least one of the collected epitope sequences were retrieved from the Uniprot database (102 for HLA-I and 71 for HLA-II epitopes) [71]. Only epitopes of these species were analyzed further. All proteomes were scanned for each epitope sequence, and epitope sequences found in only one proteome (i.e., species-specific epitopes) were kept for further analysis. Highly similar epitope sequences were identified using Clustal Omega [72] and excluded as follows. A protein distance matrix was created and epitopes were discarded iteratively. In each iteration, the epitope pairs with the lowest k-tuple distance were identified. Then, the epitope with the highest average similarity to all other sequences was excluded. Iterations were repeated until distance values less than 0.5 (corresponding to greater than approximately 50% sequence identity) were eliminated from the matrix [73]. Note that this filtering procedure was carried out separately for epitope sequences bound by HLA-I and HLA-II.

Binding affinities of the remaining 3,265 HLA-I epitope sequences to 346 HLA-A, 532 HLA-B and 225 HLA-C alleles were predicted with the NetMHCpan-4.0 algorithm [74]. The binding of 2,691 HLA-II epitope sequences to 162 HLA-DRB1 alleles was predicted using the NetMHCIIpan-3.1 algorithm [30]. All 162 alleles are present in at least one of the human populations studied here (see below). The “pep” sequence input format was used for both HLA-I and HLA-II epitope-binding prediction. A binding affinity threshold of 50 nM was applied, yielding peptides that are likely to be immunodominant [39]. For alternative binding threshold definitions, see S4 Fig. For each binding threshold, epitope-binding promiscuity was defined as the fraction of the epitope set bound by a given allele.

Calculating epitope-binding promiscuity using in vitro data

To determine the epitope-binding promiscuity of HLA-DRB1 alleles based on previously published experimental data, we used the IEDB database [38]. Specifically, we downloaded all MHC ligand and T-cell assay data, which was available for 48 HLA-DRB1 alleles. Binding data of 20 alleles screened for at least 100 ligands each were further analyzed. The epitope set of each allele was filtered for highly similar sequences, as described above. As the majority of in vitro assay data were available in a binary format (i.e., presence or absence of binding), promiscuity was calculated as the fraction of positive binding assays for a given allele.

Calculating promiscuity levels of human populations

To calculate population-level promiscuity values, we obtained HLA allele frequency data from the Allele Frequency Net Database (AFND) and the International Histocompatibility Working Group (IHWG) populations [35, 36]. Haplotype-level data of the 13th International HLA and Immunogenetics Workshop (IHIW) populations were downloaded from dbMHC (National Center for Biotechnology Information [NCBI]; ftp://ftp.ncbi.nlm.nih.gov/pub/mhc/mhc/Final%20Archive). Additionally, allele frequency data of the 14th and 16th IHIW populations, as published by Riccio and colleagues [37], and populations in the AFND were used in the analyses. To avoid potential confounding effects of recent genetic admixture and migration, we focused on native populations, similarly to previous studies (S1 Table) [7, 8]. We excluded IHWG populations reported to deviate from Hardy-Weinberg equilibrium [37]. Among the AFND populations and IHWG populations without haplotype-resolution data (14th and 16th IHIW), those comprising less than 100 genotyped individuals or those in which the sum of allele frequencies deviated from 1 by more than 1% were excluded. Populations reported in multiple databases were included only once in the analysis.

For each HLA loci, we calculated mean population promiscuity by averaging promiscuity values of alleles weighted by their relative frequencies in the populations. In all of these calculations, we used standardized (i.e., z-score) promiscuity values to make the in silico and in vitro values more easily comparable. Finally, when calculating population-level promiscuity based on in vitro promiscuity data, we excluded populations for which in vitro promiscuity values could be assigned to less than 50% cumulative allele frequency.

To tackle the issue of nonindependence of data points, we focused on populations instead of countries and grouped those populations that have highly similar HLA allele compositions, based on standard measures of genetic distance (see below). We merged populations with highly similar HLA allele compositions, allowing us to avoid pseudoreplication of data points while retaining informative allele frequency differences between populations residing in the same broad geographical areas.

To this end, we first generated a genetic distance matrix between populations with the adegenet R library using allele frequency data of the examined locus. We used the Rogers’ genetic distance measure [75] because it does not assume that allele frequency changes are driven by genetic drift only, an unlikely scenario for HLA genes. Next, populations were merged using a network-based approach. Populations were treated as nodes and two nodes were connected if their genetic distance was under a cutoff value. Populations were grouped in an iterative manner. In each iteration, all maximal cliques (i.e., subsets of nodes that are fully connected to each other) in the network were identified. Maximal cliques represent groups of populations in which all populations have similar allele compositions to each other. Then, mean genetic distance within each clique was calculated. The clique with the lowest average distance was selected and its populations were grouped together. Then, this clique was deleted from the network. Iterations were repeated until no maximal cliques remained in the network. Grouping of populations was carried out using different distance value cutoffs (1st, 5th, 10th, and 15th rank percentile of all distance values). The resulting population groups and the individual populations that remained in the network were treated as independent data points in subsequent statistical analyses. Mean promiscuity level in population groups was calculated by averaging population promiscuity values.

Unless otherwise indicated, all figures are based on population groups using the 15th percentile genetic distance cutoff value. Importantly, using different cutoffs has no impact on our results (S3 Data). Finally, we note that genetic differences among human populations mostly come from gradations in allele frequencies rather than from the presence of distinctive alleles [76]. Therefore, traditional clustering of populations based on HLA composition would have been ill-suited for our purposes.

Pathogen diversity

Data on 309 infectious diseases were collected from GIDEON [41]. For each disease, the number of causative species or genera (when species were not listed for the genus) was determined using disease information in the GIDEON database, as described previously [42]. Causative agents were classified into obligate intracellular and extracellular pathogen groups based on a previous study [7] and literature information. Putative facultative intracellular pathogens were excluded from the analysis. Diseases caused by agents that could not be clearly classified were also excluded from the analysis. Extracellular and intracellular pathogen diversity (richness) of each country was approximated by the number of identified endemic extracellular and intracellular species, respectively.

Finally, we assigned country-level measures of pathogen and HLA diversity to population groups as follows. For each population group, extracellular and intracellular pathogen counts were calculated by averaging the corresponding country-level values across the populations within the group. For example, if a population group contained two populations residing in neighboring countries, then we assigned the average pathogen diversity of the two countries to it.

Literature mining of associations between HLA alleles and infectious diseases

To examine associations between selected HLA allele groups and infectious diseases, we carried out a systematic literature search on PubMed database using the following terms:

“assoc* drb1 12 02”, “assoc* drb1 1202”, “assoc* drb1 12 01”, “assoc* drb1 1201”, “assoc* dr12”, “assoc* drb1*12”, “assoc* drb1 03 01”, “assoc* drb1 0301”, “assoc* dr3”, “assoc* drb1*03”, “assoc* dr17”, “infect* drb1 12 02”, “infect* drb1 1202”, “infect* drb1 12 01”, “infect* drb1 1201”, “infect* dr12”, “infect* drb1*12”, “infect* drb1 03 01”, “infect* drb1 0301”, “infect* dr3”, “infect* drb1*03”, and “infect* dr17”.

“assoc*” and “infect*” refer to any word beginning with these letters.

Each resulting paper containing HLA association data was examined, and statistically significant associations between allele groups (DRB1*03 or DRB1*12) or common alleles in allele groups (DRB1*12:01, DRB1*12:02, DRB1*03:01) and infectious diseases were collected. Associations with diseases caused by intracellular pathogens were excluded from the analysis. HLA-disease associations were classified as beneficial or detrimental, if all related studies supported the beneficial or detrimental role of HLA allele/allele group in the development or course of the given disease. Otherwise, the association was classified as controversial. The results were summarized (S3 Table), and statistical association between beneficial/detrimental effects and high/low promiscuity across allele groups was determined by a Fisher’s exact test.

Amino acid distance between DRB1 alleles

We used amino acid distance as a proxy for phylogenetic distance between pairs of DRB1 alleles. To this end, nucleotide sequences of DRB1 alleles that contained full exon 2 and 3 regions were downloaded from the IPD-IMGT/HLA database [19]. To limit our analyses to alleles that have an impact on the inferred promiscuity level of a population, we considered only those sequences that had a nonzero frequency in at least one human population (see above). From allele groups that code for the same protein sequence (synonymous differences, differentiated by the third set of digits in the HLA nomenclature), one of the alleles was randomly chosen. This selection procedure resulted in 96 alleles. Multiple alignment of nucleotide sequences was performed using the MUSCLE algorithm as implemented in the MEGA software [77] and converted to protein sequence alignments. Amino acid distance was calculated using the Jones-Taylor-Thornton substitution model in MEGA [77] (Fig 3A). Epitope-binding region sites—as defined previously [30]—were excluded when calculating amino acid distance. The rationale behind this exclusion is that these sites are known to be under positive selection [78, 79] and are therefore less informative on evolutionary distance. Additionally, by removing these sites, the amino acid distance remains independent of promiscuity predictions. Finally, as intragenic recombination may distort the inference of evolutionary distance, we identified such events across all alleles following the protocol of Satta and colleagues [80] using GENECONV [81] and RDP algorithms [82] as implemented in the RDP software [83]. Recombinant alleles were removed when calculating amino acid distance.

Sequence diversity and promiscuity fragility

We first defined the epitope-binding region of HLA-DRB1 alleles, as previously [30]. To estimate sequence diversity along the epitope-binding region, we employed two measures: standard Shannon entropy [84] and nucleotide diversity (π), a widely employed measure of genetic variation [85].

Using the protein sequence alignment of the 96 alleles defined above, we calculated amino acid sequence variability as the Shannon entropy of the given amino acid site as follows: where P_i is the fraction of residues of amino acid type i at a given site, and M is the number of amino acid types observed at that site.

Nonsynonymous nucleotide diversity (π_A) measures the average number of nonsynonymous nucleotide differences per nonsynonymous site between two randomly chosen protein coding DNA sequences from the same population [85, 86]. π_A was calculated for each amino acid site in the epitope-binding region for each population using DnaSP software [87] and custom-written R scripts. Nucleotide sequences of DRB1 alleles were downloaded from the IPD-IMGT/HLA database [19].

The calculation is based on the work of Nei and colleagues [85] using the equation where x_i and x_j are the frequencies of the ith and jth alleles in the population, respectively, and is the number of nonsynonymous nucleotide differences per nonsynonymous nucleotide site between the two codon sequences of the given amino acid site in the ith and jth alleles. To calculate π_A for each population, allele frequency data of human populations were obtained, as described earlier (see above). An overall nucleotide diversity index was calculated by averaging π_A across populations.

To calculate each amino acid site’s impact on epitope-binding promiscuity (promiscuity fragility), promiscuity was predicted for each 19 possible amino acid change along the epitope-binding region of each of 96 alleles. The fold difference in promiscuity resulting from each amino acid substitution was calculated. The median promiscuity fold difference of each possible allele and amino acid change combination (96 × 19) was used to estimate promiscuity fragility at each amino acid position. As some of the 19 possible amino acid changes are not accessible via a single nucleotide mutation, and the accessible amino acid changes can have different likelihoods based on the codon sequence of the site and the genetic code, we also calculated promiscuity fragility based on each nonsynonymous nucleotide substitution of the codon instead of each amino acid substitution of the site.

Statistical analysis and graphical representation

All statistical analyses were carried out in R version 3.2.0 [88]. Smooth curves were fitted using the cubic smoothing spline method [89].