Introduction

Human birth weight is under stabilizing selection, seeking a balance between extremes of high and low in order to optimize pregnancy outcomes and reduce the risk of fetal and maternal perinatal mortality [1]. Very low birth weight is associated with greater risk of stillbirth and preterm labour, and increased offspring cardiometabolic risk in later life [2–4]. Conversely when babies are born very large, the mother is at risk of experiencing prolonged obstructed labour, which may lead to haemorrhage and/or sepsis. Likewise, the child is at risk of shoulder dystocia and asphyxiation during birth, as well as cardiometabolic morbidities in later life [5,6].

Whilst mother and offspring share 50% of their nuclear genomes, they are antigenically distinct. Thus, invading fetal trophoblast cells will express molecules encoded by paternal genes, identifying the fetus as ‘non-self’ to the mother. This dynamic brings into play aspects of the immune system, namely the interaction between maternal killer immunoglobulin-like receptors (KIR) expressed on uterine natural killer cells (uNK) and human leuokocyte antigen-C (HLA-C) present on fetal trophoblast cells [7,8]. The HLA-C gene belongs to the HLA class I region (corresponding to major histocompatibility complex [MHC] class I), a highly variable region of the genome that plays a key role in antigen presentation from intracellular sources (including from invading intracellular pathogens), and recognition of one’s own cells [9]. HLA-C is the dominant ligand for KIRs in humans and is categorized into two groups: C1 and C2 based on the amino acid present at position 80 [10,11]. HLA-C binding with KIRs regulates uNK function and trophoblast invasion into the uterine wall by producing either positive or negative signals via cytokine secretion depending on whether the KIR possesses activating or inhibitory function [12–15]. KIR genotypes can be grouped into two broad haplotypes, A and B, based on their gene content. The KIR A haplotype contains fewer genes which encode mostly inhibitory receptors which bind HLA-C, while the KIR B haplotype contains numerous genes encoding mainly activating receptors, although only one, KIR2DS1, can bind with HLA-C [16,17]. Both HLA-C and KIR genes are highly polymorphic, meaning there are many possible maternal KIR and fetal HLA-C genetic combinations arising from great numbers of genotypes and considerable haplotype diversity [9,18].

Certain combinations of maternal KIR and fetal HLA-C alleles have been previously associated with birth weight. Specifically, Hiby et al. 2014 reported that the cytokine-inhibiting maternal KIR AA genotype was observed more frequently in low birth weight pregnancies, while the cytokine-activating maternal KIR BB genotype was observed more frequently in high birth weight pregnancies compared to average birth weight pregnancies [19]. When investigating the effect of the maternal KIR2DS1 locus and combinations of fetal HLA-C2 alleles on birth weight, the authors found an average increase in birth weight of ~250g in pregnancies where the mother possessed KIR2DS1 and the fetus possessed more HLA-C2 alleles than their mother. This was in comparison to pregnancies where mothers had 0 copies of KIR2DS1 and the fetus had fewer or equal numbers of C2 alleles compared with their mother. It is worth noting that the effect sizes reported in the Hiby et al. study are an order of magnitude greater than the main effect of other common variants that have been robustly associated with birth weight [19–21]. The authors also found that maternal KIR2DS1 was only associated with increased offspring birth weight when the fetal C2 allele was paternally inherited compared to maternally inherited [19]. Fetal growth and birth size are dependent on sufficient nutrient delivery via the placental spiral arteries [22,23]. These associations are consistent with KIR and HLA molecule interaction in transformation of the spiral arteries [12], i.e., combinations of fetal HLA-C2 and maternal activating genotypes (KIR BB) may excessively promote arterial transformation, enhancing fetal growth, whilst combinations of fetal C2 and maternal inhibitory genotypes (KIR AA) may insufficiently promote arterial transformation, leading to reduced fetal growth [17].

However, the sample size of the Hiby et al. study was small (n = 1,316 pregnancies), especially compared with recent genome-wide association studies (GWAS) [20,21], and though single nucleotide polymorphisms (SNPs) in the HLA and KIR gene clusters have been identified as associated with birth weight in GWAS [20,21], the key interaction findings have not been tested for replication in large independent samples to date. Whilst direct genotyping of KIR and HLA regions is costly and time intensive, statistical methods are now available to impute HLA and KIR haplotypes using SNP genotyping data [24–27] and provide an opportunity to investigate HLA/KIR interactions in much larger samples.

In this study we used statistical imputation of genome-wide SNP data to generate HLA and KIR information in five cohorts containing large numbers of mother-offspring pairs (total n = 10,602). Using the estimated KIR haplotypes and HLA-C genotypes, we tested for association between birth weight and maternal KIR/fetal HLA-C genetic combinations.

Methods

Imputation of HLA types

In the UKB, HLA alleles were imputed to a multi-population reference panel centrally by UKB using the HLA*IMP:02 [25] software. HLA genotypes were imputed at four-digit (also known as two-field) resolution for 11 classical HLA genes, indicating both the serological antigen grouping or allele family (first field, e.g., HLA-C*01) and the subtype of the allele encoding the specific HLA protein (second field, e.g., HLA-C*01:01) [52]. A quality metric (Q) is reported of the absolute posterior probability of the allele inference for each genotype call [29]. We then used Q to determine a hard call for each HLA genotype to use in our association analyses. For EFSOCH, ALSPAC, and BiB, HLA alleles were imputed locally using SNP2HLA [26] software from which the dosages for the HLA-C alleles were extracted for analysis. As the BiB data were originally genotyped on the two different arrays, imputation was performed for each group separately and then merged into one dataset prior to association analyses. We imputed HLA data for the HAPO study, using a multi-ethnic HLA reference panel and the Michigan imputation server (Minimac4) [27], applied to extracted genotype data from 26,000,000–34,000,000 bp (human genome build 37) on chromosome six. After HLA data for each cohort was imputed and the dosage data or Q-values were extracted specific to the necessary HLA-C alleles, all dosages/Q-values were converted to hard calls using set thresholds as follows: 0 = dosage ≤ 0.3; 1 = dosage ≥ 0.7 and ≤1.3; 2 = dosage ≥1.7, otherwise set to missing. For association analyses, we derived HLA-C1 or -C2 group alleles according to their allele families based on the amino acid at position 80, with the C1 ligand carrying asparagine (HLA-C*01, 03, 07, 08, 12, 14, 16 alleles), and C2 carrying lysine (HLA-C*02, 04, 05, 06, 15, 17, 18 alleles) in alignment with previously published studies [10,11].

A summary of all genotype and HLA imputation software used, genotype imputation reference panels used, amount of KIR region SNP overlap with the KIR*IMP reference panel, final number of SNPs used for KIR imputation, and KIR imputation sample sizes are found in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Overview of genotyping, HLA, and KIR imputation materials.

https://doi.org/10.1371/journal.pgen.1012102.t001

Association between maternal KIR types, fetal HLA-C, and offspring birth weight

We performed multivariable mixed linear regression to explore the association between offspring birth weight and maternal KIR A vs B haplotypes (AA, AB/BA, BB genotypes) as well as the amount of KIR2DS1 (0, 1, 2 copies of the gene) in the mothers under 3 different conditions of fetal HLA: in the presence of fetal HLA-C2 (coded as 0 or 1 for C2 group allele), in the presence of more HLA-C2 alleles in the fetus relative to the mother (i.e. mC1C1/fC1C2 and mC1C2/fC2C2), and when fetal HLA-C2 is inferred to be inherited from the father (i.e. mC1C1/fC1C2); models are described in Table 2. These groupings of subjects and the combinations analyzed were defined in accordance with those previously investigated by Hiby et al. 2014 [19]. For inferred inheritance of fetal HLA-C2 from the father, mother-child pairs where both mother and offspring were heterozygous (mC1C2/fC1C2) were not included in the analysis as paternal origin could not be determined unambiguously. Mixed linear regression models were used to account for mothers with multiple offspring in the sample, with mother’s ID set as a random effect. Our study sample included only singleton pregnancies from live births carried to term (≥37 weeks and <43 weeks) according to gestational age data where available (information on gestational age is unavailable in UKB). For each cohort prior to modelling, we used its sample size to estimate the number of standard deviations within which most birth weight observations are expected to fall (EFSOCH = 3, HAPO = 4, ALSPAC = 4, BiB = 4 SDs, respectively). Birth weight observations outside of this range were classified as outliers and excluded prior to analysis (observations excluded: EFSOCH = 0, HAPO = 474, ALSPAC = 138, BiB = 5). For UKB specifically, where there were individuals with birth weight self-reported at multiple assessment centre visits, we took the mean of the multiple values, otherwise taking the single visit where they had data. If there was a difference in absolute value of multiple reported birth weights ≥1 kg, the individual was excluded. Birth weights <2.5 kg or >4.5 kg were also excluded from analysis for the UKB sample to reduce bias from possible reporting errors and to ensure, as far as possible, that preterm births were removed as information on gestational age was not available. This left final mother-offspring pair sample sizes for performing association analyses as follows: UKB = 3,165, EFSOCH = 625, HAPO = 470, ASLPAC = 4,550, and BiB = 1,792 (Table 3). All models were adjusted for offspring sex, gestational age (except UKB), genotyping batch, the first five genome-wide principal components (PCs) in the mother and offspring, and year of birth (UKB only). All analyses were conducted separately in each cohort before performing a fixed effect, inverse-variance weighted meta-analysis to estimate the interaction effects between maternal KIR and fetal HLA on birth weight across the entire study population. Characteristics of each population sample are described in Table 3.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Models of the associations tested between maternal KIR types, fetal HLA-C2, and offspring birth weight.

https://doi.org/10.1371/journal.pgen.1012102.t002

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Descriptive characteristics for participants in each cohort.

https://doi.org/10.1371/journal.pgen.1012102.t003

Results

Quality of KIR imputation

The KIR types across the five study populations were well imputed, achieving both high estimated imputation accuracy and posterior probabilities overall. Imputation accuracy was estimated at over 95% for both the KIR2DS1 locus and the A vs. B haplotype, with the exception of imputation of the A vs. B haplotype in the ALSPAC and BiB cohorts where accuracy was 81% accuracy (Table 4). The ALSPAC and BiB genotype data was more limited in its ability to be imputed due to the coverage of this region on the genotyping chip and thus did not achieve as high a broad haplotype estimation but still provided reasonable confidence at an accuracy over 80%.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Estimated KIR imputation accuracy across all cohorts for the KIR A vs. B haplotype and KIR2DS1 locus.

https://doi.org/10.1371/journal.pgen.1012102.t004

Little evidence of association between offspring birth weight and maternal KIR types or fetal HLA-C2

After meta-analysing the maternal KIR and fetal HLA effects on birth weight across the five cohorts of mother-offspring pairs, no interaction effect was observed between either the maternal A vs. B haplotype or the maternal KIR2DS1 locus and fetal HLA-C2 when C2 was present in the fetus (yes vs. no; models 1 & 2), when the fetus had more C2 compared to the mother (models 3 & 4), and when fetal C2 was paternal in origin (models 5 & 6). Under each condition of fetal HLA-C2, no change in offspring birth weight was found for each additional maternal B allele or KIR2DS1 allele, with nearly all effect estimates close to the null with wide confidence intervals (Figs 1 and 2, respectively). When performing a sensitivity analysis removing the UKB cohort, as the only sample with birth weight not adjusted for gestational age, the results were unchanged (see S1 and S2 Figs).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Meta-analysis of the estimated change in offspring birth weight (g) for each additional maternal KIR B allele.

(A) In the presence of fetal HLA-C2 (n = 10,591). (B) In the presence of more HLA-C2 alleles in the fetus relative to the mother (n = 10,562). (C) When fetal HLA-C2 is paternal in origin (n = 10,589).

https://doi.org/10.1371/journal.pgen.1012102.g001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Meta-analysis of the estimated change in offspring birth weight (g) for each additional maternal KIR2DS1 allele.

(A) In the presence of fetal HLA-C2 (n = 10,591). (B) In the presence of more HLA-C2 alleles in the fetus relative to the mother (n = 10,562). (C) When fetal HLA-C2 is paternal in origin (n = 10,589).

https://doi.org/10.1371/journal.pgen.1012102.g002

Discussion

In our investigation combining data from five different study populations, we found little evidence that maternal KIR haplotype was associated with a change in offspring birth weight in the presence of fetal HLA-C2. There were also no associations observed with birth weight when examining the dosage of the activating receptor-encoding KIR2DS1 gene, when the fetus possessed more HLA-C2 than the mother, or when the fetal HLA-C2 was paternal in origin. This contrasts with the findings of Hiby et al. 2014 [19] where the presence of maternal KIR2DS1 combined with fetal C2 of paternal origin resulted in an average increase in birth weight of ∼250 g (total n = 1,316 pregnancies) compared to offspring without maternal KIR2DS1 and less or equal fetal C2; we observed a null effect in a sample of approximately 10,600 pairs, which was an order of magnitude larger than that of Hiby et al.

We aimed to replicate these associations but were unable to find evidence of an association between HLA and KIR combinations with birth weight. Rather than test presence vs. absence of maternal KIR2DS1, we elected to code the variable as gene copy number as output by the KIR imputation software; however, we would not expect this to alter the ability to detect an association with birth weight. We additionally tested HLA-C in combination with a maternal KIR A vs B haplotype as this is representative of the inhibitory vs activating genotype. We would expect this to reflect a similar pattern of association as with the KIR2DS1 locus on its own, perhaps with some attenuation as KIR2DS1 may not be present on all individuals but similarly found no association in this variation of analysis. Some key differences to note between these two studies include the fact that we imputed both KIR and HLA types in our study, whereas Hiby et al. [19] were able to directly type mothers and babies in their smaller sample. Our imputation estimated KIR and HLA alleles with high accuracy, but there will be some level of uncertainty in the genotype calls made for individuals; any genotyping errors will bias results towards the null. We also removed birth weight outliers on both ends of the spectrum, keeping our sample to uncomplicated pregnancies without extreme birth weight babies. Hiby et al. [19], though also investigating uncomplicated pregnancies, investigated birth weights >5^th centile with opportunity to include high birth weights if clinically determined to have no known medical cause, therefore including a greater range of birth weights than our study. Even so, in our included measures of birth weight covering normal pregnancy, we should expect to see the same previously observed patterns of association if the underlying biological mechanisms are acting as proposed. Given the null findings in our much larger sample, exclusion of outliers in our analyses alone is not sufficient to explain the lack of an association if a true effect of 250g higher birth weight in maternal KIR2DS1 carriers and fetal carriers of paternal HLA-C2 exists.

Previous work has been done to determine modelling methods that most accurately capture maternal-fetal genotype (MFG) interactions, such as that by Clark et al. 2016 and 2017, which developed the extended quantitative-MFG (QMFG) test, a linear mixed model that can apply familial multi-locus genotype data [53,54]. Their model was designed to address the issues of low power or incorrect conclusions obtained when using standard models that only consider offspring genotypes [53]. In the Clark et al. 2017 study, they used this model to the investigate the same relationship between maternal KIR2DS1, more fetal HLA-C2 than the mother, and birth weight as in Hiby et al. 2014, using the same sample of white European individuals (n ≈ 1,300 mother-child pairs) [54]. They expanded this replication by specifically testing an interaction between the loci and if there are additional independent effects of each locus. They found that the model that best fit this data and assessed these relationships, models the genetic effects of maternal KIR2DS1, more HLA-C2 in the fetus, and an interaction between them (denoted as Model 1 in their study). With this model, they found a significant KIR–HLA-C interaction effect on birth weight that explained ∼1% of the phenotypic variation in birth weight [54]. The model they used is essentially Model 4 in our study presented in Table 2 and Fig 2b: a linear mixed model testing the same genetic effects with both loci included as well as an interaction term. With respect to covariate adjustment, we adjusted our tested models for offspring sex and gestational age (except UKB) as they did, with the added benefit of adjustment for mother and child principal components in our models to guard against effects of population stratification. We have thus used a method previously determined to be the most appropriate to assess the maternal-fetal genetic relationships at these loci but were unable to replicate an effect in our analysis. It is important to reiterate that although Clark et al. 2017 were able to replicate an association between maternal KIR2DS1, greater HLA-C2 in the fetus, and increased birth weight as in the Hiby et al. 2014 study, they used the same small sample of mother-child pairs while our study increased sample size by approximately a factor of 10. Therefore, regardless of model similarities or differences, the impact of sample size on results must be considered.

The most likely explanation for our lack of evidence of a relationship between maternal KIR and fetal HLA-C on birth weight is that the previous association was a false-positive result, driven partly by the relatively small sample. It is well established that the lower a study’s power, the more likely a positive finding is to be false [55]. To our knowledge, Hiby et al. [19] also did not control for population substructure as we have done here with the inclusion of ancestry PCs. Unaccounted for differences in ancestry, even within a sample restricted to a single continent of origin, is a common reason for false-positive findings due to allele frequency differences between population subgroups among cases and controls [56]. The effect sizes observed previously [19] are not realistic for common complex traits and inconsistent with the effect sizes observed for common variants for birth weight and other traits that have been investigated over the years with GWAS [21,57,58]. Though our confidence intervals are wide, they are small enough to rule out the effect size found by Hiby et al. [19].

It is possible that the published associations of KIR and HLA-C2 are true with respect to pregnancy complications and the extremes of birth weight. Those pregnancies would be much more likely to be influenced by placental insufficiency or dysfunction which could be what these genes and their molecules truly influence rather than weight of the fetus per se. Though strides have been made with respect to GWAS and identification of genetic variants influencing pregnancy belonging to the mother, fetus, or shared by both [20,21,59–61], there is still much work to be done in the exploration of maternal-fetal interaction and how genes work together to influence or cause birth and healthy aging outcomes. New and emerging large scale genetic studies identifying variants associated with phenotypes such as pre-eclampsia [62] and placental weight [63], may offer novel avenues of investigation in terms of testable instruments and sizeable population samples to draw from. It is also worth noting that a recent study able to partition maternal and fetal effects of classical HLA alleles on birth weight found only a nominal association (P < 0.05) of HLA-C on birth weight while certain HLA-A and HLA-B alleles showed a stronger association (P < 0.001) with birth weight after Bonferroni correction for multiple testing [61]. Evaluation of immunogenetic relationships would additionally benefit from the use of methods which could type individuals at these challenging loci even more accurately on a larger scale. The rise of high throughput whole genome sequencing technologies makes the availability of such data likely in coming years.

We are limited in our ability to generalize these findings to non-European populations as we did not use an ancestrally diverse set of individuals. Some additional limitations exist within the KIR imputation software itself and how it applies the data to be imputed. The estimated accuracy of the KIR loci is calculated based on the level of intersection between the reference panel and the input data [24]. It is highly dependent on how many of the 301 reference SNPs overlap with those in the study population sample, with 12 of these SNPs deemed “highly informative”. The accuracy figures in Table 4 are an estimate of how well each KIR type can expect to be imputed using SNP genotype data coming from the same population as the reference panel. The accuracy in actuality may be worse if the underlying ancestry of the population sample differs from that of the reference panel or if the genotype quality is poor. This is reflected in how well the distributions of the posterior probabilities of the most likely allele at each KIR locus align between the reference panel and the input dataset, for which our results largely showed very good concordance – an indicator that the KIR loci are imputed well, though less so for KIR A vs. B in the ALSPAC and BiB cohorts likely due to poor SNP coverage of the genotyping chip in this region (see S3–S8 Figs). Adding further support for our study findings relative to our imputation quality, we performed a sensitivity analysis using simulations to examine the effect of varying imputation accuracy on the power to detect a main effect of maternal KIR2DS1 as well as an interaction between maternal KIR2DS1 and fetal HLA-C2 on offspring birth weight. The results indicated that we maintain very high power to detect main effects and interactions in our study, even with substantially reduced KIR imputation accuracy (simulated as low as 70%; see S1 Text). This can be attributed to having a sample size an order of magnitude larger than the original Hiby et al. study. Lastly, we sought to replicate initial published findings by Hiby et al. [19] tested in a population of uncomplicated pregnancies; we did not test these same associations in complicated pregnancies with a greater spectrum of extremes, e.g., with very low or high birth weights and pre-eclampsia, where associations could be more evident if they reflect the known underlying biological mechanisms.

Overall, though observational studies provide evidence for the role of maternal KIR and fetal HLA-C2 in pregnancy outcomes impacted by placental dysfunction, and by extension birth weight, we were unable to detect evidence of an association of these receptor-ligand pairs with offspring birth weight. Our findings here reinforce the importance of replication and use of the largest sample sizes available to maximize power, as not all observed genetic associations may be valid, even with strong biological plausibility. More research is needed to validate this relationship, or lack thereof, in more diverse populations and in complicated pregnancy cases in similarly robust sample sizes. Having successfully demonstrated the use of widely available SNP genotype data from large population studies to estimate individuals’ KIR and HLA genotypes, new statistical methods along with growing biobanks and consortia should be harnessed to better inform robust conclusions regarding maternal-fetal genetic interplay.

Supporting information

Acknowledgments

This research has been conducted using the UK Biobank Resource under application number 53641.

This study represents independent research supported by the National Institute of Health Research (NIHR) Exeter Clinical Research Facility. The views expressed are those of the author(s) and not necessarily those of the NHS, the NIHR or the Department of Health and Social care.

The Exeter Family Study of Childhood Health (EFSOCH) was supported by South West NHS Research and Development, Exeter NHS Research and Development, the Darlington Trust and the Peninsula NIHR Clinical Research Facility at the University of Exeter. The opinions given in this paper do not necessarily represent those of NIHR, the NHS or the Department of Health. We would like to acknowledge Andrew Hattersley as the principal investigator, and Bea Knight for her contribution to data collection, of the EFSOCH study.

We are extremely grateful to all the families who took part in this study, the midwives for their help in recruiting them, and the whole ALSPAC team, which includes interviewers, computer and laboratory technicians, clerical workers, research scientists, volunteers, managers, receptionists and nurses. This publication is the work of the authors and C.S.D and R.M.F will serve as guarantors for the contents of this paper.

Born in Bradford is only possible because of the enthusiasm and commitment of the children and parents in BiB. We are grateful to all the participants, health professionals, schools and researchers who have made Born in Bradford happen.

This project utilised high-performance computing funded by the UK Medical Research Council (MRC) Clinical Research Infrastructure Initiative (award number MR/M008924/1).