Signatures of pathogen-driven selection and Austronesian gene flow of Papua New Guinea HLA alleles

Summary

Human leukocyte antigen (HLA) class I and II are cell surface proteins that display peptide antigens to immune cells, thereby mediating detection of infected cells and production of antibodies. Pathogen exposure and demographic events, including local adaptation and admixture, have driven and maintained exceptional polymorphism of HLA genes across human populations. Papua New Guinea has a complex demography, with geographically distinct populations in the highlands and lowlands and exceptional linguistic heterogeneity throughout the island. The lowland populations retain signatures of Austronesian expansion ∼3,000 years ago. Papua New Guinea populations are also differentially exposed to endemic malarial pathogens, with a greater burden in the lowlands. We analyzed genome-wide autosomal SNP data together with HLA allele sequences, linguistic, and geographical data from 337 Papuans. We find the substructure of HLA alleles to be highly correlated with altitude in Papua New Guinea, a signal that is distinct from the rest of the genome. In addition, specific HLA-B and HLA-DP alleles in lowland groups have a greater number of homozygous genotypes than expected under neutrality. Some of these HLA alleles are of Austronesian genetic ancestry. We find that the HLA-binding repertoires at candidate loci are significantly enriched for antigenic P. falciparum-derived peptides. Together, these results indicate that pathogen-driven selective pressures correlate with the observed HLA genetic substructure in Papua New Guinea, highlighting the critical importance of characterizing highly complex HLA variation in understanding differences in disease susceptibility across diverse human groups.

Polymorphism of human leukocyte antigens (HLA) guides differential immune responses to infection. We analyze how HLA evolved across the distinct environments of Papua New Guinea. We find that HLA diversity correlates with altitude and that naturally selected variants may confer protection against malaria in the lowland region.

Introduction

Polymorphism of the human leukocyte antigen (HLA) genes, encoded in the major histocompatibility complex (MHC), represents an exceptional example of diversifying adaptation, driven by immune response to varied pathogens.^1,2,3,4,5,6 HLA class I (HLA-A, -B, and -C) presents intracellular-derived peptides at the cell surface, informing cytotoxic CD8⁺ T cells of ongoing infections through interaction with somatically evolved T cell receptors.^7,8,9 HLA class II (HLA-DR, -DQ, and -DP) can present pathogen-derived extracellular antigens to CD4⁺ T cells to stimulate antibody production.^10,11,12 Natural selection diversifies the peptide binding repertoires of HLA, strongly influencing specific pathogen resistance or susceptibility.^13,14 A subset of HLA class I allotypes also has one of the sequence motifs (named A3/A11, Bw4, C1, or C2) that permit interaction with natural killer (NK) cells through binding killer cell immunoglobulin-like receptors (KIRs).^15,16,17,18 In addition, some HLA class II molecules can bind specific receptors that modulate NK cell and T cell activity.^19,20,21 Thus, HLA class I and II both have critical roles in innate and adaptive immune responses.

Papua New Guineans show remarkably high levels of variation in terms of language,^22,23 altitude of residence, and pathogen exposure.^{24,25,26,27,28,29} Consequently, genetic substructure is more pronounced than that observed in other regions of similar size.^30,31 Papua New Guinea consists of four administrative regions, namely the Momase, Southern, Islands, and Highlands, the latter having an average altitude of 1,600 m. There are more than 800 languages spoken in Papua New Guinea, classified into 40 distinct families, including Trans-New Guinean (TNG), Austronesian (AN), and Sepik-Ramu (SR).^32,33 The first arrival of modern humans in Southeast Asia, including present-day Papua New Guinea, occurred around 50,000 years ago.^34,35 Studies of genome-wide array data, whole-genome sequences, and mitochondrial DNA show a divergence of lowland and highland groups around 10,000–20,000 years ago, coinciding with an in situ Neolithic transition.^30,36 After this initial split, language played an important role in the gene flow and migration events among the groups.^{22,30,35,36,37} Substructure within HLA class I has also been identified, especially regarding highlands and coastal population groups,³⁸ where linguistic divisions do not completely account for the observed allele frequency differences.³⁹ Genetic ancestry in Papua New Guinea was also influenced around 3,000 years ago by the AN expansion.^40,41,42 This migration is linked to the spread of AN languages across Oceania.^{40,41,43,44,45} The genetic impact of the AN expansion is supported by mitochondrial DNA³⁶ and Y chromosome haplogroups^46,47 and by the HLA genes, where the presence of HLA class I and II alleles having an East Asian origin has been detected in multiple Oceanian locations, including Papua New Guinea.^{48,49,50,51,52}

Malaria is endemic in Papua New Guinea,^53,54 where its burden is the highest across Asia and the Pacific,⁵⁵ and Plasmodium falciparum is the predominant pathogen.^{25,56,57,58,59} Malaria incidence and transmission vary substantially.^60,61,62,63 Since mosquito survival is restricted by altitude, malaria incidence is greatest in the lowlands, where it is hyperendemic in the East Sepik province.^61,64,65,66 Signals of positive selection in lowland groups, including the presence of the Duffy antigen FY^∗A^null allele,⁶⁷ a specific complement receptor 1 variant,⁶⁸ and high frequency of α-thalassemia,⁶⁹ have been attributed to protection from malaria.^70,71 Consistent with the rapid evolution of this pathogen, differential associations of specific HLA allotypes with protection or susceptibility are observed across affected populations.^72,73,74,75

The aim of this study is to determine the genetic substructure of HLA within Papua New Guinea and examine how this diversity has evolved. Previous studies used low-resolution genotyping of HLA alleles,^{38,39,48,50,51} and these were analyzed independently from the remainder of the genome, which hindered any ability to disentangle the action of natural selection from that of neutral factors. Here, we analyzed a dataset that covers a wide range of sampling locations and linguistic affiliations, comprising 645 individuals, 337 of whom also have overlapping genome-wide array data.⁵² The subject of the previous study was a KIR allele having distinctive ancestral and functional properties.⁵² In the present study, through combining autosomal SNPs with high-resolution HLA data, we find that altitude is more correlated with genetic variation in the MHC region than in the rest of the genome. We next identify specific high-frequency HLA class I and II alleles having an excess of homozygous genotypes in lowland groups, potentially related to protection against malaria, and show that some of these alleles arrived in Papua New Guinea with the AN expansion.

Subjects, material, and methods

Samples and data

We analyzed the HLA class I and II alleles of 645 Papua New Guineans.⁵² Of these 645 individuals, 337 individuals had overlapping genome-wide SNP data (described below). We annotated the individuals using a set of curated non-genetic variables: language family (TNG, AN, SR, and mix), geographic region (Highlands, Southern, Momase, and Islands), and altitude (highlands, lowlands, and islands)^30,76 (Figure 1A; Table S1). For example, Momase is a region of northeastern Papua New Guinea that borders the coast and contains populations speaking all three major language families (Figure 1A). The Sepik-speaking groups sampled here live in a lowland area. Assignment of population groups was based on the sampling location and affiliations of the individuals and their parents.

Figure 1.

Papua New Guinea dataset

(A) Distribution of the sampled groups. Shapes indicate language group (TNG, Trans-New Guinean; AN, Austronesian; SR, Sepik-Ramu; MIX, mixture), solid lines indicate lowlands, and dashed lines indicate highlands.

(B) Principal-component analysis (PCA) using genome-wide SNP data. Each point represents an individual, with colors and shapes the same as in (A). See Bergström et al.³⁰ for a more detailed description of the fine-scale genetic substructure.

(C) Summary of AMOVA results; see Table S4 for extended results.

The samples were sourced from a library established in the 1980s through a series of collaborative projects between the Institute for Medical Research in Papua New Guinea, Liverpool School of Tropical Medicine, and the University of Oxford, UK. The samples were collected, following informed consent, to understand genetic differences between highland and lowland populations in relation to differential disease risk. All samples were collected at the time with regulatory approval from the Papua New Guinea Institute of Medical Research and the Papua New Guinea Medical Research Advisory Council and were endorsed by the Public Health Department. Updated approval for contemporary genetic analysis was provided by the Institute for Medical Research in Papua New Guinea and the Oxford Tropical Research Ethics Committee (OxTREC). Additional approval to analyze the data was granted by the Colorado Multiple Institutional Review Board (COMIRB).

We sequenced the HLA class I and II alleles of the 645 individuals using an established capture protocol as previously described.⁷⁷ Briefly, for each sample, 300 ng genomic DNA was fragmented into 900-bp pieces, and libraries were prepared using the Kapa Hyper Prep protocol (Kapa Biosystems, Wilmington, MA) with the following modifications. A dual size selection was performed after the post-ligation cleanup. In the first size selection, 35 μL of Ampure beads (Beckman Coulter, Brea, CA) plus 45 μL of H₂O were added to 50 μL of the sample. In the second size selection, 15 μL of Ampure beads were added to 125 μL of the sample. In the library amplification step, 7 cycles were performed to achieve optimal library yield. To label the library obtained from each DNA sample, we used 96 unique dual-index Illumina adaptors that were synthesized by Integrated DNA Technologies (Coralville, IA). Libraries were pooled in batches of 96 and then subjected to enrichment for the complete set of HLA genes using a pool of oligonucleotide probes manufactured as previously described.⁷⁷ The enrichment was performed using the Nextera Rapid Capture Exome enrichment protocol (Illumina, San Diego, CA). Here, the hybridization was incubated at 95°C for 5 min, at 70°C for 4 min, cooled to 58°C over the course of 8 h, then maintained at 58°C for 10 h. Sequencing was performed on a MiSeq instrument (Illumina) using V3 chemistry, and the sequencing read length was 2 × 300 bp. The capture and sequencing protocol has been validated across multiple population samples.¹⁸ HLA alleles were determined from the sequence data using NGSengine 1.7.0 (GenDX, Utrecht, the Netherlands). The sequence reads are available at https://doi.org/10.7910/DVN/JA0U1Z. HLA haplotypes were phased using the expectation-maximization (EM) algorithm from Arlequin v.3.5.2.⁷⁸

We included Multi-Ethnic Genotyping Array (MEGA) data from Papua New Guineans obtained from two datasets: 312 individuals from Bergström et al.³⁰ and 25 individuals from the Oceanian Genome Variation Project (OGVP).⁷⁹ After applying the same quality control (QC) filters as previously,⁵² the dataset has 471,417 autosomal SNPs. For those analyses requiring unlinked markers, we performed linkage disequilibrium (LD) pruning using PLINK1.9⁸⁰ (window size of 200 SNPs, shift of 50 SNPs, and r² of 0.5), resulting in 247,344 unlinked SNPs. In total, we analyzed concordant SNP data, HLA genotypes, and altitude-level classification for 337 individuals, along with geographic region and language affiliation data from a subset (n = 300).

For reference groups, we included the Indigenous Taiwanese from the OGVP⁷⁹ and three 1000G phase 3 populations: Northern Europeans from Utah (CEU), Han Chinese in Beijing, China (CHB), and Kinh in Ho Chi Minh City, Vietnam (KHV),⁸¹ which were used in the population substructure analyses and local ancestry inference (see below) to add context in the structure results and as a reference for the AN-related ancestry. After merging and QC filtering (genotype missingness < 10% and allele frequency > 1%), the complete dataset contains 463,750 autosomal SNPs and was jointly phased using SHAPEIT v.4.2.2.⁸² Using HIBAG v.1.36.4 with the pre-trained Asian model that included 720 samples,⁸³ HLA alleles for the OGVP panel were previously imputed with an accuracy of 96% compared to the results obtained using targeted sequencing.⁵² HLA data for the 1000G samples were obtained from a published study,⁸⁴ which has HLA calls for A, B, C, DPB1, and DRB1 genes generated from sequence data.

Genetic substructure analyses in Papua New Guinea

A principal-component analysis (PCA) was performed using the smartpca program from EIGENSOFT 8.0.0⁸⁵ on the LD-pruned genome-wide SNP dataset described above. For HLA, a PCA was performed on allele frequencies by population group rather than individual genotypes (see details below). Analysis of molecular variance (AMOVA) statistics were computed for HLA class I and II data using Arlequin v.3.5.2^78,86 and for genome-wide SNP data using the poppr v.2.9.4 R package.⁸⁷ To identify the major determinants of substructure, we tested groupings that include combinations of altitude, geographic region, and language affiliation. We next performed a PCA on the HLA allele frequencies using the prcomp function of the R base package stats⁸⁸ with default arguments, following a published approach⁸⁹ and including those alleles with frequency > 1.5%. We first plotted the PC coordinates for the population groups (region-language and altitude-level groups) using the R package factoextra (https://github.com/kassambara/factoextra). We next computed PC coordinates for the HLA alleles using the get_pca_var() function from factoextra, which uses the PC loadings (i.e., prcomp return value called “rotation”) and the standard deviations of the PCs (i.e., prcomp return value called “sdev”). Then, PC allele coordinates were superimposed onto the PC coordinates of the population groups (region-language and altitude-level groups) to help visualize how strongly each HLA allele influences the PCA results.

HLA fixation indices (F_ST) were calculated using Arlequin v.3.5.2,⁷⁸ and F_ST for the LD-pruned SNP data were calculated using EIGENSOFT 6.0.1.⁸⁵ Multidimensional scaling (MDS) plots were generated using the cdmscale function from R v.4.3.1. Groups with fewer than 5 individuals were excluded from the analyses to avoid biases due to sample size differences. Differences between PCA and MDS analysis are expected, as PCA captures the maximum variation in the data and its components are orthogonal, whereas MDS maintains pairwise distances between groups.

Detecting HLA alleles under natural selection

We performed the Ewens-Watterson test for homozygosity for each HLA gene using pypop 1.2.0,⁹⁰ which handles multi-allelic data and implements a Markov chain Monte Carlo method to obtain the null distribution and report the p value. We also used chi-squared tests to determine any Hardy-Weinberg equilibrium (HWE) deviations in the observed proportion of HLA homozygotes of each allele, using R v.4.3.1.⁸⁸ We further explored those alleles with a significant HWE deviation (p < 0.05), a frequency higher than 20%, present in at least 5 individuals, and with more than 1 homozygote. We examined whether the increased homozygosity observed in HLA is also present across the genome (i.e., an indication that high genetic drift, not positive selection on HLA, causes the signal). Here, we used the genome-wide array data to compute the SNP nucleotide diversity (π) in 100-bp windows, using VCFtools v.0.1.15⁹¹ and runs of homozygosity (ROHs) segments using PLINK1.9.⁸⁰ PLINK was used with the default parameters and the following exceptions: a maximum gap of 100 kb and a minimum ROH length of 500 kb and 50 SNPs, as previously applied for array data.⁹²

Extended haplotype homozygosity (EHH) was calculated using selscan v.1.2.0,⁹³ using the genome-wide array data of individuals with the HLA allele of interest. EHH represents the probability that two randomly chosen chromosomes that carry the core haplotype are homozygous at all SNPs for the entire interval.⁹⁴ In the present study, the core haplotype is the one carrying the HLA allele of interest. We identified the core locus as the position at which all individuals carrying the HLA of interest have the same SNP allele, and EHH estimates for the reference or alternative haplotype were selected based on the core locus for that HLA allele.

As an independent assessment of the positive selection signal, we computed the integrated haplotype score (iHS).⁹⁵ Here, we used the lowland Papuan whole-genome sequences from the Simons Genome Diversity Project (SGDP),⁹⁶ since the number of SNPs from the genome-wide array that are within the HLA genes and have ancestral and derived state information is not sufficient for the iHS test. We first imputed the HLA-B and DPB1 alleles in the 7 lowland SGDP individuals using the pre-trained Asian model of HIBAG v.1.36.4⁸³ and identified 1 HLA-B^∗13:01 homozygote and 3 heterozygotes, 4 DPB1^∗04:01 homozygotes and 1 heterozygote, and 1 DPB1^∗05:01 heterozygote. These genotype frequencies are consistent with the frequencies in the dataset from the present study (Table S2). However, given the low SGDP sample size, we only have power to examine DPB1^∗04:01. Since iHS calculation relies on ancestral and derived allele assignment for haplotype construction, we used the ancestral-state annotation from the 1000G phase 3 data.⁸¹ Selscan v.1.2.0⁹³ with default parameters was used to compute the iHSs and normalize them across 100 frequency bins. The observed selection signal is due to DPB1^∗04:01, as all SNPs within the negative iHS peak (found within the DPA1 and DPB1 genomic region) have ancestral alleles in DPB1^∗04:01⁺ individuals.

Inferring local ancestry for the HLA allele haplotypes

RFMix v.2.03⁹⁷ was used to infer the local ancestry of chromosome 6 in Papuan individuals, using a two-reference panel that includes 178 East Asians (proxy for the AN component⁹⁸) and 177 Papuan highlanders. Regarding the construction of the reference panels, we included individuals based on the global ancestry proportions previously estimated⁵²: (1) East Asian AN (Indigenous Taiwanese) and non-AN-speaking (CHB and KHV) unrelated individuals who have 5% or less non-East Asian global ancestry and (2) Papuan highland unrelated individuals who have 0.01% or less East-Asian ancestry. RFMix was then run using default parameters, with the following exceptions: 3 EM iterations, 85 generations since the admixture event,³⁰ and a minimum of 5 reference haplotypes per node.⁹⁸ To assign the ancestry of the HLA alleles of interest, we next identified the haplotype (maternal or paternal) from the array data that corresponds to the HLA allele and filtered the RFMix inference using a posterior probability of 0.95.

Nomenclature

We refer to any distinct DNA sequence that spans a coding region as an allele and any allele that encodes a distinct protein sequence as an allotype. The HLA-DPA1 and DPB1 genes encode the DPα1 and DPβ1 proteins, respectively.⁹⁹ We use HLA-DPA1 or DPB1 when referring to alleles and DPα1 or DPβ1 when referring to the respective allotypes.

KIR ligands

HLA class I alleles were classified into groups based on their capacity for encoding KIR ligands: A3/A11, Bw4, C1, C2, or none (a list of allotypes and binding motifs can be found in Pollock et al.¹⁸).

HLA peptide binding repertoire prediction

Peptide binding prediction was performed using NetMHCpan-4.1 and NetMHCIIpan-4.0,¹⁰⁰ using a previously compiled dataset of 232 antigenic proteins from 27 key human pathogens (2–27 proteins per pathogen, with a mean of 9).^3,101 Each protein sequence was broken into peptides of 9 amino acid (aa) residues for HLA class I and 15 residues for HLA class II. Peptides predicted to be strong binders were determined as those having rank scores of <0.1% for HLA class I and <2% for HLA class II, which are the default thresholds for NetMHCpan-4.1 and NetMHCIIpan-4.0.¹⁰⁰ The distinct thresholds correspond to distinct peptide binding properties, where HLA class I binds short peptides tightly and HLA class II binds longer peptides more loosely.¹⁰⁰

For each HLA allotype we detected in Papua New Guinea (n = 48 HLA class I; n = 71 HLA class II), we counted the number of pathogen-derived peptides predicted to bind strongly. For each given pathogen, we plotted the number of peptides per HLA allotype to assess the percentile rank distribution of a given HLA allotype. For comparison across a representation of global populations, we performed the same test using every HLA-B or DPα1-DPβ1 allotype present in the Solberg et al. dataset (497 populations, 66,800 individuals).¹

Antigenicity estimation

Using Antigenic Protein and Peptide Ranker (APRANK), we next estimated antigenicity of all peptides that NetMHCpan-4.1 and NetMHCIIpan-4.0 predicted to bind HLA strongly and are derived from any of the P. falciparum strains/isolates. This method integrates a curated set of molecular features (predictors) to obtain a combined protein antigenicity score (range: 0–1) and minimizes variability in antigenicity estimations obtained when using single predictors.¹⁰² As a positive control, the authors of APRANK tested their method by comparing predicted antigenicity of P. falciparum peptides with known seroprevalence scores, showing a positive correlation.¹⁰² From the list of predictors that APRANK integrates, we used (1) presence of signal peptide cleavage sites with SignalP v.5.0¹⁰³; (2) three dimensional structure disorder, using Iupred3¹⁰⁴; (3) presence of transmembrane helices, using TMHMM v.2.0¹⁰⁵; (4) presence of O-glycosylation sites with NetOglyc v.4.0¹⁰⁶; (5) secondary structure and aa relative surface exposure, using NetSurfP v.3.0¹⁰⁷; (6) presence and localization of tandem repeats, using XSTREAM v.1.73¹⁰⁸; (7) aa molecular weight and isoelectric point, using Pepstats from EMBOSS v.6.6¹⁰⁹; and (8) cross-reactivity with human host proteins and self-similarity (see Ricci et al.¹⁰² for a detailed explanation on how each predictor can be useful to estimate peptide antigenicity). Given the endemicity of P. falciparum in Papua New Guinea, we note that the previously published pathogen set³ only included proteins from the isolate 3D7, which is a clone derived from the West African NF54.¹¹⁰ Thus, we also incorporated orthologous proteins derived from Asian and Southeast Asian strains having available NCBI genome annotations: NF135 (Cambodia),¹¹¹ CAMP (Malaysia),¹¹² Dd2 (Indochina),¹¹³ and K1 (Thailand).¹¹⁴

In addition to the APRANK antigenicity score, we considered seroprevalence, defined as the percentage of subjects in an independent study of individuals exposed to P. falciparum who produced antibodies that bind the protein from which a given peptide was derived.¹¹⁵

From the NetPan-predicted binding repertoires of HLA allotypes showing evidence of positive selection, we identified a set of P. falciparum-derived peptides with the greatest potential to generate an immune response. A given high-binding peptide was required to pass each of the following three thresholds: high protein antigenicity (APRANK score ≥ 0.60), high CD4 or CD8 immunogenicity scores (≥50 for CD4 or >0 for CD8; generated using previously developed tools¹⁰¹ available at http://tools.iedb.org/main/tcell/), and conservation across 2 or more P. falciparum strains/isolates. A given peptide was also included if it had been experimentally validated as a T cell epitope. Here, we used the search tool in the Immune Epitope Database for published in vitro assays: https://www.iedb.org/.¹⁰¹

Results

HLA genetic variation correlates significantly with altitude

In this study, we compared genome-wide autosomal SNP data with HLA class I and II genotypes from 337 unrelated individuals spanning the four municipal regions of Papua New Guinea (Figure 1A; Table S1). A PCA generated from genome-wide SNP data showed that genetic variation in Papua New Guinea is structured by geography and language (Figure 1B), as previously indicated.³⁰ We detected 9 distinct alleles for HLA-A, 22 for HLA-B, 17 for HLA-C, 6 for HLA-DPA1, 13 for HLA-DPB1, 14 for HLA-DQA1, and 23 for HLA-DRB1 (Table S2). After phasing the alleles, we observed 102 distinct HLA class I haplotypes (Table S3).

Using AMOVA, we first examined which non-genetic factors correlate with genetic diversity within Papua New Guinea. We used a set of established variables³⁰: language family (TNG, AN, SR, and mix), municipal region (Highlands, Southern, Momase, and Islands), and altitude (highlands and lowlands; Figure 1A). The major factor characterizing genetic substructure for both autosomal SNPs and HLA alleles in Papua New Guinea is language family, where variation among groups represents 2.85% and 9.16% of the total, respectively (Figure 1C; Table S4). This finding suggests that language facilitates gene flow among groups in Papua New Guinea, consistent with previous observations.³⁰ Interestingly, altitude is a significant determinant of among-group variation for HLA (5.4%, p < 10⁻⁵) but not for the autosomal genome-wide data (0.003%, p = 0.315; Figure 1C; Table S4), which could indicate HLA-specific selective pressures driven by differences in pathogen transmission and abundance across highlands and lowlands.^60,61,62,63

Showing among-group variation of 2.28% genome wide and 7.51% for HLA alleles, respectively, the “geographic region and language group” combination is a significant factor differentiating Papua New Guinea populations (Figure 1C; Table S4). However, geographic region alone explains only 1.65% and 5.5% of among-group variation (Table S4). We further explored geographic and linguistic differentiation by computing genetic distances using F_ST (Table S5) and visualized them using MDS. As expected, when analyzing genome-wide SNPs or HLA alleles, MDS differentiated the study groups, including three reference populations: CEU, CHB, and KHV (Figure 2, top). We then considered population structure within Papua New Guinea. In analyzing genome-wide data, TNG- and SR-speaking groups were distinct from AN-speaking groups (Figure 2, bottom left). Although MDS performed using genome-wide data supports stratification by language along dimension 1, this nuance was not observed when analyzing HLA. For HLA, the first dimension suggested that Momase Sepik-speaking and Island groups are distinct from the remainder of the populations (Figure 2, bottom right). The MDS computed from F_ST distances of SNPs throughout the MHC genomic region (which contains the HLA genes) shows closer clustering by language (Figure S1), most likely caused by multiple non-coding and intergenic variant positions present in the MHC.

Figure 2.

In Papua New Guinea, HLA genetic substructure is distinct from the rest of the genome

Multidimensional scaling (MDS) plots from F_ST for genome-wide SNP data (left) and HLA class I and II alleles (right) with Papua New Guinea groups with (A) and without (B) reference populations: CHB, Han Chinese from Beijing, China; KHV, Kinh from Ho Chi Minh City, Vietnam; CEU, Northern Europeans from Utah. Suffixes correspond to language groups (TNG, Trans-New Guinean; AN, Austronesian; SR, Sepik-Ramu).

The HLA alleles driving the PCA clustering patterns (Figure 3A) correspond to those having the most differentiated allele frequency distributions among populations (Figure 3B; Table S2). For example, DPA1^∗01:03 is informative for the Momase Sepik-speaking group because all studied individuals (n = 12) have this DPA1 allele. They have a higher proportion of A3/A11⁺HLA-A, and C2⁺HLA-C alleles (Figure S2A) and, consequently, the highest number of HLA haplotypes encoding two KIR ligands (Figure S2B; Table S3). The genetic differentiation of the Momase Sepik-speaking group is more evident for HLA than for the remainder of the genome.

Figure 3.

Specific HLA class I and II alleles drive genetic substructure

(A) Shown are PCA plots for region-language and altitude-level groups generated from HLA allele frequencies. PC coordinates (scaled loadings) for the HLA alleles are superimposed on the PCA plot to visualize alleles that contribute to the group substructure. Alleles labeled in color have PC1 or PC2 coordinates > 2 SD compared to the other alleles, shown in gray text (see subjects, material, and methods for details). Suffixes correspond to language groups (TNG, Trans-New Guinean; AN, Austronesian; SR, Sepik-Ramu).

(B) HLA class I allele frequencies for each group. Alleles are colored by KIR ligand motif: A3/A11 (yellows), Bw4 (greens), C1 (reds), C2 (blues), and grays for those lacking a motif. Only alleles with >1.5% frequency are included.

Another distinctive pattern of HLA in the MDS analysis (Figure 2) is that TNG-speaking groups form a cline based on geography (i.e., Highlands-Momase-Southern). Consistently, HLA class I allele frequency distributions follow this order (Figure 3B), especially for A^∗24:02 and C^∗01:02. Evidence for this trend is also found in the frequency of HLA class I C1⁺ KIR ligands (Figure S2A) and in the decreasing number of HLA haplotypes encoding two KIR ligands (Figure S2B; Table S3), where the groups that are further away from the central highlands resemble mainland AN-speaking groups. A PCA from HLA frequencies using altitude-based groupings (Figure 3A) indicates that there are 5 HLA class I and 5 HLA class II alleles driving the clustering, of which 3 are found at high frequencies in the highlands (HLA-A^∗24:02, -C^∗01:02, and -DPA1^∗01:03) and 7 in the lowlands (HLA-B^∗13:01, -B^∗40:02, -C^∗15:02, -DPA1^∗02:02, -DPB1:05:01, -DQA1:05:05, and -DRB1:11:01). Of these allotypes, HLA-A^∗24:02 has been identified as subject to positive selection across Oceania.⁵²

Some high-frequency HLA alleles in lowland groups have an excess of homozygosity

We next explored the hypothesis that positive selection has contributed to the distinct HLA frequency spectrum of the Momase Sepik-speaking population. We identified two HLA alleles to have a significant excess of homozygous genotypes (B^∗13:01 and the DPA1^∗01:03-DPB1^∗05:01 haplotype; Tables 1, S6, and S7; Figure 4 ). We note that HLA class II molecules are expressed as αβ heterodimers, such that the DPA1-DPB1 combination encodes a functional allotype that can be subject to natural selection.¹¹⁶ The observations of increased homozygosity at B^∗13:01 and DPA1^∗01:03-DPB1^∗05:01 are independent of each other, due to a lack of LD between them. Indeed, no individual has both B^∗13:01 and DPA1^∗01:03-DPB1^∗05:01 on the same haplotype (Table S3B). Indicating that the increased homozygosity is specific to HLA, there is no statistical difference in genome-wide mean nucleotide diversity between the Momase Sepik-speaking and the other groups (Figure S3A). In addition, the mean length in genome-wide ROH across groups is similar (Figure S3B). However, the Momase Sepik-speaking group has a higher mean sum length for the shorter ROH (1–2 cM) than other lowland groups (Figure S3B).

Table 1.

Frequencies and statistical significance from HWE (p HWE) of alleles having an excess of homozygosity

HLA allele	Group under selection	Allele frequency (%)	p HWE	HLA-DP haplotype frequency (%)
B∗13:01	Momase_SR	33.33	0.025	–
DPA1∗01:03	Momase_SR	100	N/A	DPA1∗01:03 - DPB1∗05:01 = 8.33
DPB1∗05:01	Momase_SR	8.33	<10−5	DPA1∗01:03 - DPB1∗05:01 = 8.33
DPA1∗01:03	lowlands	49.53	1.6 × 10−4	DPA1∗01:03 - DPB1∗05:01 = 2.28
DPB1∗05:01	lowlands	35.57	3.1 × 10−4	DPA1∗01:03 - DPB1∗04:01 = 27.31
DPB1∗04:01	lowlands	30.41	<10−5	DPA1∗02:02 - DPB1∗04:01=0.46
DPA1∗02:02	lowlands	40.09	<10−15	DPA1∗02:02 - DPB1∗05:01 = 31.94

DPA1-DPB1 haplotype frequencies are shown on the right. p is not applicable (N/A) when all individuals are homozygous. See Tables S2, S9, and S11.

Figure 4.

Specific HLA alleles in lowland groups show evidence of positive selection

(A) De Finetti diagram showing the proportion of observed heterozygotes (y axis) and the allele frequency for each HLA allele having increased homozygosity (see Table 1); HWE expectations are shown in a dashed line (neutrality).

(B) EHH for homozygous B^∗13:01 individuals from the Momase Sepik-speaking group (Momase_SR) (left y axis) and local ancestry assignment for the haplotypes carrying the B^∗13:01 allele (right y axis); Papuan ancestry is in green and East Asian is in brown.

(C) EHH for lowland homozygous individuals for the specified DPA1 or DPB1 alleles; each image shows an overlaid DP combination.

HLA-B and -C show high LD in all populations.¹¹⁷ Although the most frequent B^∗13:01 haplotype in the Momase Sepik-speaking group also carries C^∗03:04 (Figure S4A; Table S3A), the allele frequency of C^∗03:04 is lower than that of B^∗13:01 (Table S2). Accordingly, the EHH for B^∗13:01 homozygous individuals decays evenly on both sides of the HLA-B gene (Figure 4B). This observation suggests that the excess of genotype homozygosity and increased allele frequency is specific to HLA-B^∗13:01.

We next analyzed broader HLA differentiation across highlands and lowlands. There are three HLA-DPA1-DPB1 allele combinations showing a significant excess of homozygotes and notably high allele frequencies in lowland groups: DPA1^∗01:03-DPB1^∗04:01, DPA1^∗02:02-DPB1^∗05:01, and DPA1^∗01:03-DPB1^∗05:01 (Tables 1, S2, S8, and S9). Here, DPB1^∗04:01 exhibits LD with DPA1^∗01:03, where 96.7% of the DPB1^∗04:01 haplotypes also possess DPA1^∗01:03, and DPB1^∗05:01 exhibits LD with DPA1^∗02:02, where 89.6% of DPB1^∗05:01 haplotypes also carry DPA1^∗02:02 (Figure S4B). Consistently, the decay of EHH for DPA1 and DPB1 homozygous individuals begins outside the DPA1-DPB1 genomic region (Figure 4C). By comparison, lowland individuals have genome-wide nucleotide diversity (Figure S3A) and homozygosity (Figure S3B) similar to that observed in the other populations. These findings are consistent with signatures of positive selection in the HLA-DP region, rather than genetic drift, generating an excess of HLA homozygosity in Papua New Guinea lowland groups. In addition, we replicated this observation using an independent dataset of whole genomes from 7 lowland Papuans (SGDP), in which we find a positive selection signal at DPB1^∗04:01 (see subjects, material, and methods and Figure S3C). The presence of multiple haplotypes and the EHH decay patterns do not support a hard sweep model and point instead to a soft-sweep event or multiple, independent selection events.

Two HLA-B^∗13:01 haplotypes have distinct origins

We investigated the ancestry of the alleles having signatures of natural selection, specifically testing for evidence of AN gene flow. We grouped Indigenous Taiwanese, CHB, and KHV as an East Asian reference and used Papuan highlanders as a proxy for Papuan-related ancestry. Local ancestry assignment was performed using genome-wide array data from the lowland populations (see subjects, material, and methods, Text S1, and Figure S5 for details on parameter selection and inference accuracy). Local ancestry results suggest that the B^∗13:01 present in lowland individuals has two distinct origins: the C^∗03:04-B^∗13:01 haplotype has an AN-like ancestry, whereas the C^∗04:01-B^∗13:01 is assigned to Papuan ancestry (Table S10; Figure 4B). This dichotomy is consistent with observed HLA-C^∗03:04 and C^∗04:01 allele frequencies across East and Southeast Asia (http://pypop.org/popdata).¹

Local ancestry inference for DPA1^∗01:03-DPB1^∗04:01 suggests that this lowland haplotype is of Papua New Guinea ancestry (Table S10). The allele frequencies of DPB1^∗04:01 are also higher in Papua New Guinea than in East Asian groups.¹ There are two subsets of HLA-DPA1^∗01:03-DPB1^∗05:01 and DPA1^∗02:02-DPB1^∗05:01 haplotypes: some haplotypes in lowland individuals are inferred to be East Asian and others as Papuan (Table S10). Interestingly, DPB1^∗05:01 genotype frequencies do not deviate from HWE in any of the East Asian reference populations we used (Taiwanese: χ² = 0.619, p = 0.43; CHB: χ² = 0.416, p = 0.52; and KHV: χ² = 0.159, p = 0.69). Moreover, although frequent, the genotype frequencies of B^∗13:01 in Taiwanese do not deviate from HWE (χ² = 0.375, p = 0.54). These observations point to local adaptation specific to Papua New Guinea. In contrast to HLA class I, ancestry assignment for HLA class II haplotypes thus suggests an endemic origin or selection from regional standing variation.

The selected HLA allotypes have a high capacity to stimulate immunity to the malaria parasite

Given the greater incidence of malaria in the lowlands than the highlands of Papua New Guinea and the potential for P. falciparum to exert strong selection pressure on humans,^118,119,120 we examined whether this parasite could be responsible for the imprint of selection we observed. We first investigated whether peptides derived from P. falciparum are represented in the binding repertoires of the four HLA allotypes having evidence for selection and then whether the peptides are likely to be antigenic.

Using an established set of pathogens with historical impact on human morbidity and mortality,³ we examined the pathogen-derived peptide binding repertoires across HLA class I allotypes present in Papua New Guinea. The number of pathogen-derived peptides predicted to bind HLA-B^∗13:01 is noticeably high (above the 90^th percentile) for five of these pathogens, including P. falciparum (Figure S6A). Thus, HLA-B^∗13:01 can present a greater number of peptides derived from P. falciparum than 90% of other HLA class I allotypes present in Papua New Guinea, including the ten most frequent (Figure 5A). Moreover, HLA-B^∗13:01 remains above the 90^th percentile among HLA-B allotypes observed in a large sample of global populations (Figure S7). Together, these findings show that HLA-B^∗13:01 has an extensive, relatively high capacity to present P. falciparum peptides to immune cells. Considering this pathogen is evolving rapidly, any HLA allotype presenting a large repertoire of distinct peptides (for B^∗13:01, n = ∼400) likely optimizes the probability of stimulating pathogen-specific immune cells.¹²¹ Importantly, however, not every peptide that binds strongly to HLA will be able to initiate an immune response.¹²² We therefore examined the antigenicity of any P. falciparum-derived peptides that could bind to HLA-B^∗13:01.

Figure 5.

HLA alleles under positive selection favor peptides derived from antigenic P. falciparum proteins

(A) Density plots showing the distribution of number of predicted peptides for all the HLA allotypes in the Papua New Guinea dataset (n = 48 HLA class I; n = 71 HLA class II), derived from P. falciparum proteins (pathogens with asterisks); colors indicate percentiles, and dotted lines represent the values for the HLA alleles under positive selection (in red) and HLA alleles with a frequency higher than 10% (in black).

(B) Protein antigenicity scores for peptides derived from each P. falciparum strain/isolate; asterisks denote statistical significance from a Wilcoxon test after multiple test correction. 3D7, Dd2, CAMP/Malaysia, and NF135 are the strains tested.

(C) Protein seroprevalence from P. falciparum 3D7 (see subjects, material, and methods) for the predicted peptides; asterisks denote statistical significance from a Wilcoxon test after multiple test correction.

Using an established pipeline,¹⁰² we determined the distribution of predicted antigenicity scores for those strong-binding peptides derived from P. falciparum. Here, in addition to African 3D7, we included three Southeast Asian strains. This analysis identified three broad groups of peptides, having low, intermediate, and high antigenicity, respectively (Figure 5B). When compared to all other HLA class I allotypes present in Papua New Guinea, the P. falciparum-derived peptides binding to HLA-B^∗13:01 had greater predicted antigenicity, and this reached statistical significance for all four pathogen strains (p < 0.05 CAMP and NF135, p < 0.01 3D7, and p < 0.001 Dd2; Figure 5C). Thus, in summary, the combination of the quantity and quality of bound peptides likely renders HLA-B^∗13:01 a highly effective immune defense initiator in the context of P. falciparum infection.

We performed a similar analysis for HLA class II. Notably, of the three selected allotypes, DPα1^∗02:02-DPβ1^∗05:01 showed skewing toward P. falciparum, with the number of peptides occurring above the 90^th percentile (Figure S6B). Here, DPα1^∗02:02-DPβ1^∗05:01 fell above the 90^th percentile of all HLA class II allotypes present in Papua New Guinea (Figure 5A) and above the 95^th percentile by comparison across global populations (Figure S7). Of their bound peptide repertoires, statistically significant deviations toward high antigenicity were observed for DPα1^∗01:03-DPβ1^∗04:01 targeting CAMP/Malaysia and NF135 (p < 0.05) and for DPα1^∗02:02-DPβ1^∗05:01 targeting NF135 (p < 0.01; Figure 5C).

Next, we examined a survey of antibody profiles obtained from subjects exposed to P. falciparum.¹¹⁵ For each peptide we identified as a strong binder, we determined the proportion of subjects who produced a protective antibody targeting the protein from which it is derived. This analysis showed that a significantly greater proportion of peptides bound by HLA-B^∗13:01 are derived from proteins that promote the production of antibodies than those bound by other HLA class I allotypes (p < 0.05; Figure 5D). A similar observation was made for DPα1^∗01:03-DPβ1^∗04:01 (p < 0.01; Figure 5D) but not for DPα1^∗01:03-DPβ1^∗05:01 or DPα1^∗02:02-DPβ1^∗05:01.

To identify those P. falciparum-derived peptides most likely to induce an immune response, we adopted empirically determined thresholds based on protein antigenicity, peptide immunogenicity, and conservation scores (Figure S8A). In this analysis, we included P. falciparum strains originating from Southeast Asia in addition to the African-origin 3D7. From 4,098 peptides identified as strong binders for the candidate HLA allotypes (Table S12), 1,432 are predicted to bind B^∗13:01 and the remainder to one or more of the three HLA-DP allotypes (Figure S8B). Applying the thresholds identified 208 of the B^∗13:01 (15%) and 726 of the HLA-DP (27%) peptides as most likely to be immunogenic (Table 2), where 50 of the latter bind all three HLA-DP allotypes (Figure S7C). Interestingly, more than 50% of these curated peptides derive from proteins of the Cambodian NF135 isolate (Figure S8C), which was not the case before applying the threshold criteria (Figure S8B). Finally, we predicted HLA-binding peptides derived from P. vivax proteins and obtained similar results, suggesting that the findings are not limited to P. falciparum (Text S2).

Table 2.

Number of P. falciparum-derived peptides presented by the HLA alleles under positive selection that also pass the thresholds for antigenicity described in Figure S8

HLA allotype	Peptides derived from P. falciparum (all tested strains)
	Total (N)	Antigenic (N)
B∗13:01	1,420	208
DPα1∗01:03-DPβ1∗05:01	958	354
DPα1∗01:03-DPβ1∗04:01	1,110	236
DPα1∗02:02-DPβ1∗05:01	1,217	309

In summary, we find evidence that the four HLA allotypes that are candidates for having been subject to natural selection in Papua New Guinea have substantial capacity to present antigenic peptides derived from the Plasmodium parasite to immune cells. This relatively high quantity likely increases the possibility that one or more of the peptides are antigenic. Indeed, both predictive and empirical data indicate that these peptides have greater potential to induce an immune response than those that bind other HLA allotypes present in Papua New Guinea. This evidence is strongest for HLA-B^∗13:01 and DPα1^∗02:02-DPβ1^∗05:01, especially for Southeast Asian strains of P. falciparum. Taken together, these findings support the thesis that differences in malaria exposure contribute to the distinctive evolution of HLA class I and II in Papua New Guinea lowlanders.

Discussion

By analyzing HLA, genome-wide autosomal data, and non-genetic information from 337 Papua New Guineans, we identify geography and language as important factors shaping the observed genetic substructure. Genome-wide autosomal SNP data are clearly structured by language family, as previously observed.³⁰ Although HLA diversity is also partially delineated by language groups, we detect two unique patterns. The first is differentiation of the Momase Sepik-speaking group, and the second is a correlation with altitude. Although genome-wide SNP data also showed qualitative differentiation from highlands to lowlands, this was only significant for the HLA genes.

We propose that selective pressures contribute to the observed genetic substructure in the HLA genes. We identify one HLA-B allele (B^∗13:01) and a functional haplotype of HLA-DP (DPA1^∗01:03-DPB1^∗05:01) as having increased homozygosity in the Momase Sepik-speakers. Although genetic drift could also increase homozygosity after a long period of isolation, as previously suggested for this group,³⁸ our results do not support this hypothesis, since the remainder of the autosomal genome does not appear to have extensive ROHs. Although natural selection was originally proposed as being responsible for altitude-based HLA allele frequency differences in Papua New Guinea, this was not formally tested at the time.³⁹ Here, we identify one HLA-B allele and three DPA1-DPB1 combinations as having signals of positive selection in lowlands, suggesting differentiation through time and/or pathogen strain.¹²³ Among them, HLA-DPA1^∗02:02, DPA1^∗01:03, and DPB1^∗04:01 are found abundantly worldwide, indicating that they could associate with broad protective function against pathogen exposures.¹¹⁷ The ancestral origin of the selected HLA alleles in Papua New Guinea can be partially traced to the AN expansion, specifically C^∗03:04-B^∗13:01 and some DPB1^∗05:01 haplotypes, in accordance with previous inferences based on HLA-C and DPB1 allele frequencies.^48,49,51 Taken together, the EHH, LD, and analyses of local ancestry indicate that these alleles are present on distinct haplotype backgrounds, exhibiting signatures of positive selection but with limited evidence of a hard sweep. Similar features were identified in populations of European and East Asian descent.^124,125 Together, these findings support the notion that pathogen-driven frequency-dependent selection is a major driver of the extreme HLA diversity observed worldwide.²

We observed that HLA-B^∗13:01 has the capacity to bind a diverse repertoire of peptides derived from P. falciparum, thereby increasing the likelihood that a subset of peptides will be antigenic and stimulate immunity to this pathogen. Indeed, we showed that B^∗13:01 presents a significantly greater number of peptides that are antigenic than other extant Papua New Guinea allotypes can present. We also showed that B^∗13:01 retains the ability to combat additional pathogens. These observations could place HLA-B^∗13:01 somewhere between a “specialist” HLA molecule focused on generating highly effective immunity toward a specific pathogen and a “generalist” with a more promiscuous repertoire.¹²² In evolutionary terms, humans are unlikely to be challenged by a single pathogen in their early lifetime, and so the ability to retain a broad response is likely essential. Moreover, considering that the distinction in peptide quantity from other allotypes is not statistically significant (α > 0.05), we assume that multiple HLA allotypes can confer a degree of protection from malaria. Here, the specific alleles offering the greatest selection advantage may fluctuate over time due to the rapid evolution of the parasite, and we expect that other high-binding allotypes may instead have been able to initiate immune responses to previously common strains. This assumption is consistent with the soft-sweep mode of selection that we detected. Our results therefore likely reflect the dynamic interactions of host-pathogen coevolution.

It is estimated that P. falciparum arrived in present-day Papua New Guinea 10,000 or more years ago.^126,127 Selection for HLA-B^∗13:01 and specific HLA-DP allotypes could have begun with the standing HLA variation and the pressure exacerbated by the widespread adoption of agriculture in Papua New Guinea ∼7,000–10,000 years ago.¹²⁸ This adoption led to a rapid expansion of P. falciparum due to increased malaria transmission.^126,129 Later, the AN migration (∼3,000 years ago^40,41,42) potentially introduced additional beneficial alleles, representing an example of adaptive gene flow. These AN alleles are not under selection in the source populations (CHB, KHV, or Taiwanese). This finding is consistent with the previously observed genetic barrier between Southeast Asian and Oceanian populations at the HLA-A and -B loci, suggesting ecological differences across these regions led to distinct immune pressures.¹³⁰ It is possible that selection began during migration to the southeast, as P. falciparum was present in other regions along the route to Papua New Guinea.¹²⁶ Evidence of a similar malaria-related selection mechanism is reported for the SLC4-A1 allele in Island Southeast Asia.^131,132

As multiple studies of sub-Saharan African populations detected HLA-B^∗13:01 rarely and found no excess of homozygosity at HLA-DPA1 or DPB1,^1,133,134 the HLA alleles detected in Papua New Guinea have no apparent relationship with malaria protection in Africa. The Southeast Asian strains are genetically differentiated from the African one^135,136 and show some similarities with an eradicated European P. falciparum strain.¹³⁷ All Southeast Asian strains used in the present study are chloroquine resistant.^111,136,138 Three distinct parasite subpopulations coexist in Southeast Asia: an ancestral antimalarial drug-sensitive group, a group resistant to artemisinin-based combination treatments (ACTs) that emerged around 25 years ago, and an older group with an admixed genetic background.^136,139,140 Among the Southeast Asian strains, we found an enrichment of HLA-binding immunogenic peptides that derive from NF135. This specific P. falciparum clone was isolated from Cambodia,¹¹¹ and it is a representative of the long-standing subpopulation of admixed parasites rather than of the recent ACT-resistant group.^136,140 Multiple functional and genetic studies have focused on NF135. For example, studies with volunteers who were naive to malaria show its higher infectivity^135,141 and earlier parasitemia with higher magnitude¹⁴¹ when compared to the NF54 strain, which might translate to more severe malaria outcomes.^136,142,143 In addition, NF135 is genetically similar to the other Southeast Asian strains,^135,136 although at a finer scale, it differs from the Papuan isolates¹³⁶: the latter were collected from 2008 to 2011 and they are ACT resistant.¹⁴⁴ The NF135 strain might then exemplify the malaria parasite driving positive selection in Papua New Guinea lowlands, given that the emergence of ACT-resistant isolates is too recent^136,140 for signatures of positive selection to be detected in the human genome.

As T cells (both CD8⁺ and CD4⁺) and NK cells have important roles in protective immunity from malaria,^{145,146,147,148,149,150,151,152} we propose three mechanisms through which positive selection on specific HLA alleles could be beneficial. The first mechanism consists of CD8⁺ T cell activation mediated by antigen-presenting cells (APCs) through cross-presentation of infected hepatocytes or by CD8⁺ T cell regulation mediated by infected hepatocytes in the tissue during malaria liver stage,¹⁵³ where HLA-B^∗13:01 complexed with P. falciparum peptides interacts with the T cell receptor. The second mechanism is related to the activation of CD4⁺ T cells mediated by APCs with their HLA-DP molecules presenting P. falciparum peptides. Indeed, two peptides predicted to bind DPα1^∗01:03-DPβ1^∗05:01 and DPα1^∗02:02-DPβ1^∗05:01 have been validated as T cell epitopes: one associated with a delayed and lower rate of P. falciparum reinfection¹⁴⁶ and the other as a P. falciparum protective epitope recognized by CD4⁺ T cells.¹⁴⁷ The third mechanism involves NK cell activity, through HLA-DP acting as a ligand for the NK cell receptor NKp44¹⁹ or through the B^∗13:01 bound with specific peptides. Although B^∗13 is not known to bind KIR3DL1 in steady state,¹⁵⁴ this receptor is peptide sensitive,⁵² raising the possibility that specific pathogen-derived peptides could enable their interaction. We caution that HLA-peptide-binding analyses were predicted in silico and require functional validation.

We note that in addition to the finding for P. falciparum, B^∗13:01 showed skewed binding, in terms of the quantity of presented peptides, toward HIV, Salmonella enterica, Schistosoma mansoni, and mumps virus. Among the pathogens examined, DPα1^∗01:03-DPβ1^∗05:01 and DPα1^∗02:02-DPβ1^∗05:01 also showed skewing toward rubella virus. At the time of sampling in the 1980s,³⁰ HIV was absent, Schistosoma was and remains absent, and Salmonella was only recently endemic to Papua New Guinea, being most prevalent in the highlands.^155,156 We therefore surmise that neither of these pathogens is responsible for the selection pressure that affected HLA allotype frequencies in the lowlands. Although both rubella and mumps were endemic at the time, with low vaccination rates, any impact on survival rates is unknown.¹⁵⁷ Thus, although malaria remains a most plausible candidate, we cannot rule out that other infections, including those we did not examine,¹⁵⁸ are responsible for the signals of selection distinct to the lowlands region that we observed.

In summary, the present study describes how geography and language track with the genetic substructure and HLA-specific selective pressures in Papua New Guinea. We find that differences in P. falciparum exposure may be responsible for the distinctive evolution of HLA in lowlanders. Specifically, one HLA-B allele and three functional DPA1-DPB1 haplotypes show the footprint of positive selection, having exceptional potential for binding antigenic P. falciparum-derived peptides, and with evidence of AN-related gene flow. These results emphasize the importance of characterizing HLA variation in diverse populations, to better understand the influence of gene flow and differences in immune response and disease susceptibility across human groups.

Data and code availability

HLA allele frequencies can be found in Table S2 of the present study. Bioinformatic code can be found at https://github.com/NFontPorterias/HLA-analyses.

Acknowledgments

We thank all the participant communities for taking part in this study. This work was supported by NIH R01 AI151549 (P.J.N.) and R35GM133531 (B.M.H.).

Declaration of interests

The authors declare no competing interests.

Published: May 4, 2026