Genomic evidence for MHC disassortative mating in humans

Abstract

Although pervasive in many animal species, the evidence for major histocompatibility complex (MHC) disassortative mating in humans remains inconsistent across studies. Here, to revisit this issue, we analyse dense genotype data for 883 European and Middle Eastern couples. To distinguish MHC-specific effects from socio-cultural confounders, the pattern of relatedness between spouses in the MHC region is compared to the rest of the genome. Couples from Israel exhibit no significant pattern of relatedness across the MHC region, whereas across the genome, they are more similar than random pairs of individuals, which may reflect social homogamy and/or cousin marriages. On the other hand, couples from The Netherlands and more generally from Northern Europe are significantly more MHC-dissimilar than random pairs of individuals, and this pattern of dissimilarity is extreme when compared with the rest of the genome. Our findings support the hypothesis that the MHC influences mate choice in humans in a context-dependent way: MHC-driven preferences may exist in all populations but, in some populations, social constraints over mate choice may reduce the ability of individuals to rely on such biological cues when choosing their mates.

1. Background

High-density genomic data have recently become a useful source of information regarding the factors influencing mate choice in human populations. In particular, assortative mating signals with respect to ancestry [1,2], social status [3,4] and phenotypic traits [5] have been reported. In addition, inbreeding levels are contrasted among human populations [6–8], probably reflecting variable frequencies of consanguineous unions in these populations [9]. Non-random mating processes with respect to biological or socio-cultural factors are attracting growing interest among geneticists because they may generate cryptic genetic structures [10,11], potentially biasing kinship and heritability estimation [12] and generating false positives when performing association studies [13–17].

One of the natural candidates when looking for non-random mating processes is the major histocompatibility complex (MHC), an ancient multigene family playing a fundamental role in immunity. MHC genes code for receptors located on the surface of a variety of immune and non-immune cells. Their role is to bind self and non-self antigens and to present them to T cells, thus initiating immune responses when antigens from intracellular and extracellular pathogens or malignant cells are recognized [18–21]. Antigen presentation is a codominant trait and heterozygotes present all antigens displayed by both homozygotes, which potentially increases their resistance to pathogens [18,22].

The MHC is known for its extremely high level of polymorphism [21,23], with over 9000 class I and 2500 class II alleles in human species [24,25]. Pathogen-driven balancing selection has probably contributed to such polymorphism, as suggested by the greater MHC polymorphism in pathogen-rich regions at a worldwide scale [26]. MHC non-random mating has been proposed as a complementary mechanism potentially promoting heterozygosity in the offspring and contributing to the maintenance of MHC polymorphism at the species level [18,19]. Indeed, although the mechanisms through which MHC influences odour and mate choice remain disputed [27], a recent meta-analysis, summarizing decades of research on vertebrate species, and in particular, in mammals, birds, fishes and reptiles, supports a female preference for MHC-diverse males and a preference for MHC dissimilarity when dissimilarity is characterized across multiple loci [28]. Preference for MHC-diverse males could provide direct benefits to females (healthier mates), as well as indirect genetic benefits to the offspring (rare beneficial alleles are more likely to be carried by MHC-diverse individuals and to be transmitted to the next generation). Preference for MHC dissimilarity could provide indirect genetic benefits to the offspring through either optimization of their immune response, or avoidance of close inbreeding if MHC functions as a kin recognition system. Selection for specific MHC alleles may also exist (for example, see [29]).

However, the situation in humans and non-human primates is more controversial [30–32]. A trend for female preference for MHC diversity was reported for non-human primates, and this effect was statistically significant in humans [31]. However, no consensus was reached regarding the preference for MHC dissimilarity, with reports of preference for MHC dissimilarity, MHC similarity or no preference at all [31,32]. Such inconsistencies may result from variable study designs throughout this field of research: some studies investigated mating preferences (focusing on odour or facial cues) while other looked at mating outcomes. An additional confounding factor is the level of ethnic heterogeneity in the population sample: signals of MHC similarity preference are overrepresented in studies with ethnically heterogeneous population samples [31]. In addition, MHC non-random mating in humans may be particularly context-dependent and may operate only under certain socio-cultural contexts. In particular, social homogamy with respect to ethnicity, language, religion, socio-economic status [3,4,33], as well as alliance rules (such as cousin marriages, whose prevalence varies across populations [9,34]) may reduce the ability of individuals to rely on biological cues when choosing their mates. Consequently, we do not expect MHC and other biological factors to have the same weight on mate choice in all human populations.

The availability of genomic data, especially when spousal data have been collected, offers an unprecedented opportunity to investigate such context dependency by disentangling biological from socio-cultural factors acting on mate choice. Notably, while socio-cultural factors shape the diversity of every portion of the genome, making spouses appear more similar or dissimilar than expected by chance, non-random mate choice with respect to the MHC will only target and shape the diversity of the MHC region. Taking advantage of this rationale, Chaix et al. previously developed a methodological frame to search for MHC non-random mating signals using an outlier approach [35,36]: indeed, under the hypothesis of MHC disassortative mating, we expect the MHC dissimilarity between spouses to be extreme in comparison to random pairs of individuals, as well as extreme in comparison to the rest of the genome (which can be considered representative of the null expectation of spouse relatedness, taking into account socio-cultural factors). Using such a methodological framework, Chaix et al. previously detected a signature of MHC disassortative mating in a sample of European American origin [35]. However, this signature has been challenged by a study claiming that such signal of MHC non-random mating was driven by outliers, was lost when minor modifications were made to the analysis procedure and that opposite-sex non-spouse pairs showed a dissimilarity signal as strong as spouses [37]. In addition, this study reported that the signal of MHC non-random mating was not found when using a more recent version of the data comprising a larger number of individuals but fewer genetic markers [37]. Laurent & Chaix disagreed with these criticisms [38,39], but this debate highlighted the need for larger datasets to explore more thoroughly the occurrence of MHC non-random mating in human populations. A recent analysis of 872 European American couples concluded for an absence of significant MHC disassortative mating [40]. However, while most studies [35,37,38] used the classical definition of the MHC region spanning 3.6 Mb [23], this study focused on a less stringent definition of the MHC region spanning over 4.9 Mb [40], a choice that may have diluted the signal of non-random mating (a non-significant trend for MHC dissimilarity between spouses was reported for this extended definition of the MHC). Here, we propose to reexamine the hypothesis of MHC disassortative mating in another large sample, the GAIN-ADHD dataset [41–44] which includes 883 European and Middle Eastern couples. We use the same methodological framework as the one Chaix et al. originally proposed [35], but we also include the changes in the methodological procedure proposed by Derti et al. [37].

2. Material and methods

(a). GAIN-ADHD dataset

GAIN-ADHD is a genome-wide association study for attention deficit with hyperactivity disorder (ADHD) based on parent–child trios from the International Multisite ADHD Genetics (IMAGE) Project sampled in seven countries: Belgium, Germany, Ireland, The Netherlands, the UK, Spain and Israel. In Germany, The Netherlands, the UK and Israel, participants were recruited from two centres. More details about recruitment and clinical description of samples are given in the papers describing this dataset [41–44]. This dataset has been cleaned and quality controlled by the National Center for Biotechnology Information (NCBI) using the GAIN QA/QC software package (version 0.7.4) developed by G. Abecasis and S. Gopalakrishnan at the University of Michigan, as described by Lasky-Su et al. [43]. The cleaned dataset contains 2732 individuals (including 895 couples) and 430 055 autosomal SNPs. Among them, 1327 SNPs were located in the MHC region (positions 29 700 000–33 300 000 on chromosome 6, in build 36.2 coordinates). This definition of the MHC is based on the classical definition of the MHC provided by Horton et al. [23], starting after the MOG gene and ending before the COL11A2 gene (when reading the short arm of the chromosome 6 from telomere to centromere), and encompassing the classical class I, II and III subregions.

(b). Preliminary analyses

We excluded all children from this dataset. In addition, we removed 8418 SNPs with minor allele frequency below 5%. Because our methodology assumes that observations are independent of each other, we removed redundant SNPs by applying an LD (linkage disequilibrium) pruning procedure retaining a subset of SNPs in low LD (r² < 0.4), using the R packages gaston [45] and GPop [46]. However, similar results were found when performing our relatedness analyses on a dataset without LD pruning (results not shown). Furthermore, in order to remove closely related individuals, we computed the genetic correlation between standardized genotypes (i.e. genotypes standardized to have mean zero and variance 1) for all pairs of individuals [47], thus identifying pairs of individuals with genetic relatedness higher than 1/8 (which is the expected relatedness between first cousins). Then, we removed for each of these pairs, one of the two individuals. In total, nine individuals were removed (six individuals from Israel, one from the UK and two from The Netherlands). Finally, we performed a principal component analysis (PCA) and we further removed six additional individuals recruited in the UK that clustered with Israel samples. The dataset ready for our relatedness analyses comprises 1798 unrelated individuals (including 883 couples and 32 single individuals) and 205 019 SNPs, of which 421 SNPs were within the MHC region. Mean heterozygosity computed over these 421 MHC SNPs for each population sample are provided in electronic supplementary material, table S1.

(c). Relatedness analyses

We estimated the genetic relatedness between spouses using the Rousset relatedness coefficient [48] between spouses i and j, defined as

$$R_{ij}=\frac{Q_{ij}-\hspace{0.17em}{Q}_m}{1-Q_m},$$ 2.1

where Q_ij is the proportion of variants which are identical by state between the two spouses and Q_m is the mean proportion of variants which are identical by state on average in the population (computed by comparing all possible pairs of individuals in each population sample). At a given SNP, the proportion of identical variants for a given pair of individuals was set to 0 if both individuals were homozygous and carrying a different allele (e.g. 00 and 11), 1 if both individuals were homozygous and carrying the same allele (e.g. 00 and 00), and 0.5 in all others cases. We checked that setting this proportion to 1 (instead of 0.5) for pairs of heterozygous individuals (01 and 01) did not impact our findings. The relatedness between spouses was summarized across the MHC region and the genome by computing for each couple i and j the mean ratio R_ij over all SNPs and then averaging this ratio over all couples. In addition, to exclude the possibility of outliers influencing our findings, we checked that similar results were found when focusing on the median of the ratio R_ij over all couples.

Significance was assessed by permuting individuals among the couples 10⁵ times, attributing a new wife to each husband. We estimated a two-sided p-value defined as the proportion of permutations in which the permuted couples had more extreme relatedness than the real couples.

In addition, to control for genome-wide effects, we asked whether the MHC region was unusual relative to the rest of the genome in terms of similarity or dissimilarity between spouses. To do so, we compared the MHC to all sliding windows of 3.6 Mb (in increments of 300 kb) having at least 100 variants, not overlapping a centromere and having similar recombination rate (±10%) as the MHC (1117 windows met these criteria). Here, it is crucial to take into account the recombination rate because the MHC has particular LD structure [49] and the variance of any statistics (here the relatedness coefficient) computed throughout the genome is influenced by the local recombination rate, such that more extreme values are expected in regions of low recombination [50,51]. We estimated the proportion of such 1117 windows exhibiting more extreme relatedness coefficient between spouses than the MHC. Such a proportion corresponds to a two-sided p-value as relatedness coefficients computed for all genomic windows similar to the MHC in terms of length and recombination rate provide an empirical null distribution.

To further check whether our results were robust to changes in the methodological procedure, we re-estimated the relatedness between spouses using two alternative statistics: first, the genetic correlation C_ij between spouses i and j based on their standardized genotypes [47]

$$C_{ij}=\frac{{\sum }_{l=1}^q{Z}_{il}{Z}_{\hspace{0.17em}jl}}{q-1},$$ 2.2

where Z_il is the standardized genotype of individual i at locus l, and q is equal to the number of SNPs included in the analysis.

Second, we computed a Z-score, indicating the deviation between the mean relatedness between spouses in comparison to non-spouses and defined as below

$$Z=\frac{m_{\mathrm{spouses}}-m_{\mathrm{non}-\mathrm{spouses}}}{{\sigma }_{\mathrm{non}-\mathrm{spouses}}},$$ 2.3

where m_spouses is the mean Rousset coefficient between spouses and m_non-spouses and σ_non-spouses are the mean and the standard error of Rousset coefficient between all possible pairs of individuals except spouses.

In addition, we computed a Z-score, indicating the deviation between the mean relatedness between spouses in comparison to opposite-sex non-spouse pairs and defined as

$$Z=\frac{m_{\mathrm{spouses}}-m_{o\mathrm{pposite}-\mathrm{sex}\hspace{0.17em}\mathrm{non}-\mathrm{spouse}s}}{{\sigma }_{\mathrm{opposite}-\mathrm{sex}\hspace{0.17em}\mathrm{non}-\mathrm{spouses}}},$$ 2.4

where m_spouses is the mean Rousset coefficient between spouses and m_{opposite-sex non-spouses} and σ_{opposite-sex non-spouses} are the mean and the standard error of Rousset coefficient between all opposite-sex non-spouse pairs. Opposite-sex non-spouse pairs include all male–female pairs that are not spouse to each other.

3. Results

We investigated MHC non-random mating signal in European and Middle Eastern populations using the GAIN-ADHD dataset, which comprises genotype data for parent–child trios sampled in Belgium, Germany, Ireland, The Netherlands, the UK, Spain and Israel. Principal component analyses (electronic supplementary material, figures S1 and S2) confirmed that most samples from the same country clustered together. In addition, while samples from Israel and Spain clustered apart from other countries, no striking genetic boundary was observed between Northern European countries (Belgium, Germany, Ireland, The Netherlands and the UK). Consequently, we analysed each country independently as well as a merged dataset grouping all individuals from Northern European countries (and referred here as the Northern European population). The motivation behind such grouping was to gain power in our statistical analyses by increasing sample size, while introducing a limited amount of genetic substructure. After quality control and preliminary analyses (including an LD-pruning step to obtain a dataset of approximately independent SNPs), the dataset comprises 883 couples (table 1) and 205 019 SNPs, including 421 SNPs within the MHC region.

Table 1.

Sample size and mean Rousset relatedness coefficient (R) between spouses at the genome-wide level and at the MHC. Italic values correspond to significant p-values considering the 5% threshold.

population	sample size (number of couples)	genome-wide		MHC
		relatedness between spouses (R)	p-valuea	relatedness between spouses (R)	p-valueb	p-valuec
Northern Europe	1330 (658)	8 × 10−4	<10−5	−0.009	0.003	0
Belgium	72 (36)	8 × 10−4	0.033	−0.011	0.343	0.264
Ireland	172 (86)	5 × 10−4	0.032	−0.018	0.068	0.003
Germany	217 (108)	5 × 10−4	0.009	−0.010	0.170	0.073
Netherlands	613 (302)	4 × 10−4	10−4	−0.010	0.040	0.010
UK	256 (126)	−8 × 10−5	0.696	−0.010	0.157	0.080
Spain	139 (69)	10−4	0.734	−0.010	0.179	0.120
Israel	329 (156)	0.003	<10−5	0.003	0.575	0.625

^aTwo-sided p-value defined as the proportion of spouse permutations yielding a more extreme R than the R observed between real spouses.

^bTwo-sided p-value defined as the proportion of spouse permutations yielding a more extreme R than the R observed between real spouses.

^cTwo-sided p-value defined as the proportion of 3.6 Mb genomic windows (with similar recombination rate as the MHC) exhibiting a more extreme R between spouses than the MHC.

We measured the genetic relatedness between spouses using Rousset relatedness coefficient (R) [48] (table 1; electronic supplementary material, figure S3). In all populations, the mean relatedness coefficients between spouses for the MHC region were negative, ranging from −0.018 in Ireland to −0.009 in Northern Europe, except in Israel where R was equal to 0.003. We tested whether spouses were more MHC-similar or -dissimilar than random pairs of individuals by permuting individuals between couples (attributing a new husband randomly to each wife) (figure 1 and table 1). We observed that real spouses from The Netherlands were more MHC-dissimilar than random pairs of spouses (p = 0.040) and marginally significant results were reached for spouses from Ireland (p = 0.068). When analysing all spouses from Northern Europe together (and permuting individuals between couples regardless of their country), we observed again that spouses were significantly more MHC-dissimilar from each other than random pairs of individuals (p = 0.003). p-values in Belgium, Germany, the UK, Spain and Israel were above 0.1.

Figure 1.

Distribution of mean Rousset relatedness coefficient computed at the MHC level for real spouses (red line) and for permuted spouses (blue distribution, 10⁵ permutations).

To control for genome-wide effects, and putative socio-demographic factors influencing mate choice, we compared these observations to the pattern of genetic relatedness across the genome. Genome-wide spouses were significantly more genetically similar than random pairs of individuals in all populations (p < 0.05), except in Spain and the UK (table 1; electronic supplementary material, figure S4). In particular, spouses from Israel exhibit similarity across the genome much higher (R = 0.003, p < 10⁻⁵) than the similarity between spouses observed in the other populations (R ≤ 8 × 10⁻⁴).

To further control for genome-wide effects, we asked whether the MHC region was unusual relative to similar regions across the genome in terms of similarity or dissimilarity between spouses, by comparing the average MHC relatedness between spouses to that of all genomic windows having the same length (3.6 Mb) and similar recombination rate (±10%) as the MHC (figure 2 and table 1). Only 0.3%, 1%, 7.3% and 8% of windows exhibited a more extreme level of genetic relatedness between spouses than the MHC when focusing on Ireland, The Netherlands, Germany and the UK, respectively. Moreover, in the merged Northern European population, the MHC exhibited the strongest non-random signal among all genomic windows with similar length and recombination rate. The MHC was not statistically unusual in comparison to the rest of the genome in terms of relatedness between spouses in Belgium, Spain and Israel (table 1). Note that this approach, which compares the MHC to genomic windows with similar recombination rate, also controls for variation in the level of genetic diversity throughout the genome since recombination and genetic diversity are correlated to each other [52]. In particular, the number of windows having a similar (±10%) recombination rate to that of the MHC (1117) is very close to the number of windows having a similar (±10%) recombination rate and heterozygosity to those of the MHC (from 1055 to 1113), confirming that the recombination rate is a good proxy for the genetic diversity. In other words, it is unlikely that the high diversity of the MHC biases our findings. To further confirm this, we compared the MHC to regions of the genome having similar recombination rate and similar level of heterozygosity (±10%) as the MHC and we replicated our findings: no window had more extreme relatedness coefficient between spouses than the MHC in Northern Europe and only 0.3 and 1% of the windows had more extreme relatedness between spouses than the MHC in Ireland and The Netherlands.

Figure 2.

Mean Rousset relatedness coefficient between spouses across 3.6 Mb sliding windows (in increments of 300 kb) throughout the genome, plotted against the window recombination rate. The red point corresponds to the MHC.

In addition, we checked that our results were robust to changes in the analysis procedure, as proposed by a study [37] that criticized the original statistical approach of the study by Chaix et al. [35]. These changes were as follows. (i) We performed a slight modification in the way the Rousset coefficient was computed by setting the proportion of identical variants between pairs of heterozygous individuals (01 and 01) to 1, instead of 0.5. (ii) We checked that our findings were not driven by outliers by performing all analyses with the median (rather than mean) Rousset relatedness coefficient between spouses. In addition, we considered alternative statistics assessing relatedness between spouses, that is (iii) the mean genetic correlation between spouses' standardized genotypes [47] and (iv) a Z-score indicating the deviation in genetic relatedness between spouses and non-spouses. These four changes in the analysis procedure had limited effect on our results, and strong evidence for MHC non-random mating was still found in Northern Europe and The Netherlands and occasionally in Ireland, as well as in Belgium, Germany and the UK (electronic supplementary material, tables S2–S5). In addition, the signals of MHC non-random mate choice in Northern Europe remain significant after multiple testing correction (the significance threshold with the Bonferroni correction is 0.006 and p-values are below 0.003 for Northern Europe).

Last but not least, we asked whether opposite-sex non-spouse pairs exhibit a similar signal of MHC dissimilarity as the spouses, as suggested by Derti et al. [37]. This hypothesis was not supported by the data since trends for MHC similarity were observed between opposite-sex non-spouse pairs in most population samples (p > 0.1) except in Germany, where a significant signal of dissimilarity was found (R = −9 × 10⁻⁴, p = 0.03) (table 2). In addition, we computed a Z-score indicating the deviation in genetic relatedness between spouses and opposite-sex non-spouse pairs. This analysis confirmed again that spouses were extreme in terms of MHC dissimilarity in comparison to opposite-sex non-spouse pairs in Northern Europe (p < 0.01) and marginally so in Ireland and The Netherlands (p < 0.06) (table 2).

Table 2.

Mean Rousset relatedness coefficient (R) between opposite-sex non-spouses and Z-score indicating the deviation of Rousset relatedness coefficient between spouses and opposite-sex non-spouses at the MHC. Italic values correspond to significant p-values considering the 5% threshold.

population	opposite-sex non-spouses		spouses versus opposite-sex non-spouses
	relatedness (R)	p-valuea	Z-score	p-valueb	p-valuec
Northern Europe	10−4	0.470	−0.102	0.003	9 × 10−4
Belgium	3 × 10−5	0.975	−0.159	0.308	0.241
Ireland	5 × 10−4	0.372	−0.173	0.055	0.052
Germany	−9 × 10−4	0.030	−0.100	0.248	0.241
Netherlands	10−4	0.148	−0.098	0.057	0.051
UK	−3 × 10−4	0.991	−0.123	0.124	0.062
Spain	10−4	0.863	−0.149	0.157	0.079
Israel	0.002	0.178	0.021	0.727	0.657

^aTwo-sided p-value defined as the proportion of 3.6 Mb genomic windows (with similar recombination rate as the MHC) exhibiting a more extreme R between opposite-sex non-spouses than the MHC.

^bTwo-sided p-value defined as the proportion of spouse permutations yielding a more extreme Z-score than the Z-score observed for real spouses.

^cTwo-sided p-value defined as the proportion of 3.6 Mb genomic windows (with similar recombination rate as the MHC) exhibiting a more extreme Z-score than the MHC.

4. Discussion

In this study, we found a signal of MHC non-random mating among Dutch couples, as well as in a merged population sample of Northern European origin: not only these couples were significantly more MHC-dissimilar than random pairs of individuals, but this pattern of dissimilarity was extreme when compared to the rest of the genome, both globally and when broken into windows having the same length and similar recombination rate as the MHC. This signal of non-random mating in Northern Europe was robust to multiple testing correction, and our observations were globally robust to small changes in the analysis procedure, particularly modifications in the relatedness measures suggested by the study [37] that criticized the original statistical approach proposed by Chaix et al. [35]. In addition, spouses from Ireland, but also from Belgium, Germany, the UK and Spain, exhibit tendencies for MHC dissimilarity, which became occasionally significant when performing these changes in the relatedness measures. This suggests a putative lack of power in these population samples to reach significance thresholds. To test this hypothesis, we performed a post hoc power analysis, evaluating the influence of sample size on the ability to detect a level of MHC non-random mating similar to the one observed among the 302 Dutch couples (electronic supplementary material, note and table S6). This analysis shows that this power is 5.5%, 9.1%, 12%, 12.9% and 17.3% when the number of couples drops to 36, 69, 86, 108 and 126, as this is the case for population samples from Belgium, Spain, Ireland, Germany and the UK, respectively. This sharp decrease in statistical power with sample size may explain why the trend for MHC dissimilarity between spouses from Ireland, the UK, Belgium and Spain was not significant.

In addition, we confirmed that such significant MHC dissimilarity signals in The Netherlands and in the merged Northern European sample were specific to spouses and cannot be generalized to all opposite-sex pairs as suggested by Derti et al. [37]. Indeed, positive MHC relatedness coefficients were observed among opposite-sex non-spouse pairs in most population samples, except in the German population sample, where opposite-sex non-spouse pairs exhibited a significant signal of MHC dissimilarity compared to the rest of genome. Although this would warrant further analysis, we hypothesize that such signal in Germany may relate to differences between males and females MHC composition that could derive from sex-specific post-copulatory selective pressures or from sexually antagonistic selection as reported in several species [53–57].

Because the dissimilarity among Dutch and more generally Northern European couples was MHC-specific and not found at the genome-wide level, it may result from MHC disassortative mating in these populations, possibly triggered by our olfactory capacity to discriminate MHC-mediated odours [18,22,27]. We explored the pattern of spouse dissimilarity throughout the MHC using a sliding-window approach and observed that the genetic dissimilarity among spouses was distributed all along the MHC, rather than localized to particular parts of the MHC, which suggests that such disassortative mating might be driven by a summation of effects over multiple genes (electronic supplementary material, figure S5). This is in line with previous analyses that reported negative results at the level of individual SNPs, HLA alleles and amino acids [35,40] and with previous observations from vertebrates species [28], which reported significant female choice for dissimilarity only when dissimilarity was characterized across multiple loci. In addition, figure S5 in the electronic supplementary material shows that the signal of non-random mating drops quickly when moving away from the classical MHC region, potentially explaining why a recent study of a large European American dataset focusing on a less stringent definition of the MHC region spanning over 4.9 Mb reported a non-significant trend for MHC dissimilarity between these spouses [40].

All couples from this study were selected for having offspring, one of them being affected by the ADHD [41–43]. It is unlikely that the inclusion criteria regarding ADHD is affecting our findings since no association was reported between MHC and ADHD [41,58]. In addition, even in the case of an association between the MHC and ADHD, this may create a bias towards MHC similarity (rather than dissimilarity) among parents of children affected by the ADHD, a pattern found in a study of spouses having a child affected by multiple sclerosis [59], a disease strongly associated with MHC [60]. Consequently, we believe that our signal is indicative of MHC disassortative mating in the general population. However, it is possible that the pattern reported here is restricted to fertile spouses, rather than couples in general, thus reflecting post-copulatory rather than pre-copulatory pressures [61–63]. Such possibility was also discussed in a previous study by Chaix et al. reporting MHC disassortative mating among spouses of European American origin (having at least a child) [35]. Nevertheless, MHC disassortative mating was also found among Hutterite couples [64], in which all couples were included, regardless of whether they had a child or not (C. Ober 2008, personal communication).

The benefits of MHC disassortative mating are still debated. Such process could either increase or optimize the level of MHC heterozygosity in the offspring, thus enhancing their resistance to multiple pathogens. Alternatively, the MHC could function as a kin recognition system to minimize inbreeding at the genome-wide level [18,22,65]. Under this latter hypothesis, a correlation should exist between spouse relatedness at the MHC and genome-wide levels. However, no such correlation was found in Ireland, The Netherlands and more general Northern Europe samples (electronic supplementary material, figure S6 and table S7). This supports the hypothesis that in this dataset, MHC disassortative mating pattern aims at enhancing the immune resistance of offspring rather than limiting their inbreeding.

When looking at the genome-wide pattern of relatedness among spouses, we found a signal of similarity between spouses in all samples except the UK and Spain. Such a pattern complements previous observations [2] and may result from the tendency to mate with someone from the same location or neighbourhood [66,67] (likely to be genetically closer than a more distant individual [68,69]), or from the tendency to mate with a distant cousin [9]. Alternatively, such a pattern may be generated under random mating if the data were collected from different geographical areas and then aggregated for the analysis. However, no genetic stratification was detected for all population samples except for Israel (electronic supplementary material, figures S1 and S2), which supports the hypothesis of geographical homogamy (tendency to mate with a ‘neighbour’) and/or distant cousin marriages as the mating process generating such assortative patterns in the European samples.

The Israeli population sample exhibits an assortative signal much stronger than the ones seen in the other population samples. In addition, we detected a specific genetic structure in this sample, with one clear genetic cluster, as well as two other less well-defined clusters, and a number of admixed individuals in between them (electronic supplementary material, figures S1 and S2). The clusters are not geographically based since they mixed individuals sampled in Tel Aviv and Jerusalem (electronic supplementary material, figure S2). In order to further understand such structures, we performed a PCA analysis of the Israel sample merged with European and Middle Eastern populations samples from the HGDP dataset [70] (electronic supplementary material, figure S7). We observed that the two less well-defined clusters fall in between European and Middle Eastern populations, as observed in previous studies of Jewish populations [71]. However, the more defined cluster distinguishes itself from all European and Middle Eastern populations, suggesting a strong genetic drift effect for this cluster. Such genetic structures may relate to the fact that the Israel samples analysed here include both Ashkenazi and non-Ashkenazi (from Arab countries) Jews (R. P. Ebstein 2017, personal communication). We hypothesize that such social stratification constrains mate choice and that the observed signal of genome-wide assortative mating reflects social homogamy [33], possibly combined with cousin marriages, a frequent practice in this country [9]. On the other hand, in Israel, the MHC exhibits no departure from the null hypothesis of random mating (table 1). This is true also when replicating the analysis in each of the three genetic clusters seen on the PCA analysis (electronic supplementary material, table S8). These observations support the hypothesis that whenever social rules over mate choice are strong, individuals are less prone to rely on biological cues, and particularly on the MHC, when choosing a mate. Similar finding was reported for the Yoruba population, in which Chaix et al. found a signal of genome-wide assortative mating possibly reflecting cousin marriages, but no significant signal of MHC non-random mating [35].

Finally, to estimate the biological importance of the signal of MHC non-random mating detected in The Netherlands and in Northern Europe, as well as its consequences in terms of offspring genetic diversity, we used the female choice model proposed by Hedrick [65]. Assuming a two-allele model, we estimated the strength of female choice (relative preference of MHC-dissimilar males over MHC similar males) to be around 7%. The decrease in homozygosity in offspring was estimated to be around 1% (electronic supplementary material, table S9). These estimates, based on this simple model, suggest that the weight of the MHC on mate choice in these populations is significant but weak, and may easily be overwhelmed by socio-cultural factors. This may explain why this effect was not found in the Israel population sample, nor in other populations such as the Yoruba [35] and South Amerindians [72], where the role played by socio-cultural factors on mate choice is important. A decrease of 1% in offspring homozygosity seems weak too and it is difficult to assess whether this has any consequence on the fitness of the offspring compared to offspring born from random pairings. However, even if the effect on offspring is low in contemporary populations, this does not mean there is no MHC non-random mating process in these populations. The signal we detect may be the heritage of millions of years of evolution, still acting to some extent in human populations, with no particular benefit for the offspring.

To conclude, our findings support the view that the MHC influences mate choice in humans in a context-dependent way [32]. In particular, the role played by the MHC seems to depend on the socio-cultural context. Indeed, we found evidence for MHC disassortative mating in Dutch and more generally Northern European human populations, but not in the Israeli sample, where a strong genome-wide signature of assortative mating was observed, probably reflecting social constraints over mate choice, such as social homogamy and/or cousin marriages. From these findings, we hypothesize that the MHC may influence mate choice in human populations only in the context of weak socio-cultural constraints over mate choice. To test this hypothesis, this approach should be extended to more population samples from all continents, taking into account the variability of human social organizations [34]. This would help assessing the respective weights of the biological and socio-cultural factors influencing mate choice in human populations. In addition, because the MHC is associated with a wide spectrum of complex diseases [73], the consequences of such MHC non-random mate choice over kinship estimation and association studies remain an open question.

Ethics

This study is based on data from the GAIN-ADHD study. Ethical approval for the study was obtained from National Institute of Health registered ethical review boards for each centre (subjects were recruited from 12 specialist centres in 8 countries). Detailed information sheets were provided and informed consent obtained from the majority of children and from all of their parents.

Data accessibility

The GAIN-ADHD dataset was obtained from the database of Genotypes and Phenotypes (dbGaP) found at http://www.ncbi.nlm.nih.gov/gap, through dbGaP accession number 34465-5. Codes are available as supplementary material. R package GPop is available in GitHub (https://github.com/genostats/GPop).

Authors' contributions

R.C., B.T., R.L. and C.D.-R. designed the study; C.D.-R., R.L., I.D. and R.C. analysed the data; B.T. implemented the female choice model; R.C. and C.D.-R. wrote the manuscript.

Competing interests

The authors declare no competing interests.

Funding

Funding support for the International Multisite ADHD Genetics (IMAGE) project was provided by NIH grants R01MH62873 and R01MH081803 to S.V. Faraone and the genotyping of samples was provided through the Genetic Association Information Network (GAIN). C.D.-R. was supported by a postdoctoral grant from the French Museum National d'Histoire Naturelle.