Cross-ancestry Genome-wide Association Studies of Sex Hormone Concentrations in Pre- and Postmenopausal Women

Abstract

Objective

Concentrations of circulating sex hormones have been associated with a variety of diseases in women and are strongly influenced by menopausal status. We investigated the genetic architectures of circulating concentrations of estradiol, testosterone, and SHBG by menopausal status in women of European and African ancestry.

Methods

Using data on 229 966 women from the UK Biobank, we conducted genome-wide association studies (GWASs) of circulating concentrations of estradiol, testosterone, and SHBG in premenopausal and postmenopausal women. We tested for evidence of heterogeneity of genetic effects by menopausal status and genetic ancestry. We conducted gene-based enrichment analyses to identify tissues in which genes with GWAS-enriched signals were expressed.

Results

We identified 4 loci (5q35.2, 12q14.3, 19q13.42, 20p12.3) that were associated with detectable concentrations of estradiol in both pre- and postmenopausal women of European ancestry. Heterogeneity analysis identified 1 locus for testosterone (7q22.1) in the CYP3A7 gene and 1 locus that was strongly associated with concentrations of SHBG in premenopausal women only (10q15.1) near the AKR1C4 gene. Gene-based analysis of testosterone revealed evidence of enrichment of GWAS signals in genes expressed in adipose tissue for postmenopausal women. We did not find any evidence of ancestry-specific genetic effects for concentrations of estradiol, testosterone, or SHBG.

Conclusions

We identified specific loci that showed genome-wide significant evidence of heterogeneity by menopausal status for testosterone and SHBG. We also observed support for a more prominent role of genetic variants located near genes expressed in adipose tissue in determining testosterone concentrations among postmenopausal women.

Concentrations of testosterone and estrogens have been implicated for their roles in the development of numerous diseases and disease-related traits, including cancers, cardiovascular disease, diabetes, osteoporosis, and metabolic syndromes (1-3). Sex-specific determinants of circulating hormone concentrations as well as sex-specific effects of hormones on disease risk have been widely recognized (4, 5). In women, concentrations of circulating testosterone and estrogen decline with age, with markedly lower natural production of estrogen with the loss of ovarian follicular activity because of menopause (6, 7).

Leveraging the release of biomarker and genome-wide genotype data in the UK Biobank (8), recent work conducted by Ruth and colleagues has provided significant insight into the genetic architecture underlying variation in circulating sex hormones, mainly in men of European ancestry (9). Hundreds of independent genetic loci have been associated with concentrations of testosterone and SHBG, 2 highly heritable phenotypes (h² = 13%-57% and 20%-68%, respectively) (9-11). However, important gaps remain for understanding the potential impact of menopause on genetic determinants of circulating concentrations of testosterone and SHBG for women.

To an even greater extent than testosterone and SHBG, the genetic architecture of endogenous estradiol concentrations in women is poorly understood. Estrogens are involved in sex-specific processes and have been associated with numerous sex-linked disease processes including breast cancer development and osteoporosis (12, 13). Detectable concentrations of estradiol in women are predominantly determined by menopausal status, with 97% of postmenopausal women in the UK Biobank having concentrations below the detectable threshold compared with 35% of premenopausal women. Consequently, an agnostic genome-wide association study (GWAS) ignoring menopausal status will be biased toward detecting loci associated with age at onset of menopause rather than true variation in estradiol (9).

The number of known associations between common genetic variants and human phenotypes continues to grow dramatically, in part because of large-scale biobanks such as the UK Biobank (8). However, genetic association studies still tend to be conducted exclusively among individuals of European ancestry (14). The implications of this lack of diversity could result in misidentification of causal variants and overestimation of effects across populations (15).

In this study, we conducted GWASs of circulating concentrations of estradiol, testosterone, and SHBG, stratified by menopausal status in women of European and African ancestry. We estimated genetic correlations across menopausal-stratified sex hormone phenotypes and explored the consistency in the genetic architecture of circulating concentrations in sex hormones across European and African ancestries.

Methods

UK Biobank Participants and Biomarkers

Details regarding survey, genotyping, and biomarker data collection have been presented elsewhere (8). Briefly, the UK Biobank study collected 34 biomarkers at baseline for ~500 000 participants along with written informed consent and extensive questionnaire information (8). Our analyses focus on measured concentrations of estradiol (field 30800-0.0), testosterone (field 30850-0.0), and SHBG (field 20830-0.0), which were measured from blood samples using, respectively, 2-step competitive analysis, 1-step competitive analysis, and 2-step sandwich immunoassay analysis on the Beckman Coulter Unicel Dxl 800. We excluded women who were pregnant at the time of data collection (field 3140-0.0). In postmenopausal-specific analyses related to estradiol, we further excluded women who had reported ever having taken hormone replacement therapy (field 2814-0.0) because exogenous hormones would likely overshadow genetic determinants of circulating estradiol concentrations and reduce our ability to find genetic loci of significance. We also excluded participants who requested to withdraw from the UK Biobank as of August 20, 2020. This research was approved under UK Biobank application 55120.

Defining Menopausal Status

We defined premenopausal status as having answered “no” to self-reported menopause (field 2724-0.0) and being younger than age 60 at the time of reference (field 21003-0.0) (7). Women who were pregnant at the time of reference were excluded (3140-0.0). Postmenopausal women were classified as having answered “yes” to having entered menopause (field 2724-0.0) and at least 2 years since their last period (field 3581-0.0), which was required to have occurred after the age of 40 years to exclude premature menopause. We excluded women who had ever used hormone replacement therapy in the postmenopausal specific analyses (field 2814-0.0).

Genetic Data and Empirically Assigning Ancestry

We use the “v3” release of the UK Biobank genetic data imputed using the Haplotype Reference Consortium and 1000 Genome Project (1KGP) as reference panels. To assign every individual in the UK Biobank to an overall ancestry group (or “superpopulation” defined as African, European, South Asian, East Asian, or Admixed American), we used the 1KGP, which has genetic data on 26 different populations worldwide as reference population, as previously described by Petersen and colleagues (16, 17). We first used Plink 1.9 (18) to identify a set of independent genetic variants using the following options (--indep-pairwise 10000 10 0.1) in the UK Biobank data. We then extracted those variants (when available) from the 1KGP data, excluding any variants with a minor allele frequency (MAF) < 0.05. To ensure that the genotype data from UK Biobank and 1KGP were aligned, we used GenotypeHarmonizer (19) to check for strand flips in the UK Biobank genotype data using the 1KGP as a reference. We then calculated principal components for both UK Biobank and 1KGP (20) on the set of extracted variants.

We use the 26 population descriptions in 1KGP to estimate within-population standard deviations for Mahalanobis distances based on the first 10 principal components and removed samples greater than 4 SDs away from the mean within their population. We then calculated Mahalanobis distances for all samples in the UK Biobank using the mean and covariance matrix from 1KGP with outliers removed. We assigned all samples to the population with the smallest Mahalanobis distance and their respective superpopulation (17). We assigned 431,239 samples as European ancestry and 9501 samples as African ancestry. We used Plink2.0 to calculate MAF, Hardy-Weinberg equilibrium, and variant missingness within each empirically assigned ancestry population.

Statistical Analyses

Genome-wide Association Studies

We used PLINK2.0 (21, 22) to perform GWAS analyses. We used GCTA (23) to identify unrelated individuals within each assigned ancestry population (--king-cutoff 0.125). We modeled all phenotypes using inverse-rank normalized transformations adjusting for age at reference (field 21003-0.0), center (field 54-0.0), genotyping array (field 22000-0.0), and the first 10 principal components (field 22009-0.1-10) as covariates within empirically assigned ancestry populations and within menopausal status for menopausal status-specific GWAS. Time since last period, created using age at last period (field 3581-0.0) and age at reference, was also included in the postmenopausal analyses to limit any bias for detecting menopause-related loci while limiting genomic inflation. We used QQ-plots to assess the presence of systematic bias. Body mass index (field 21001-0.0) was additionally included in all SHBG analyses to increase comparability with previous GWAS findings (9) to identify variants independent of their effect on body mass index; this has been shown to increase statistical power by reducing trait variance. We filtered variants with Hardy-Weinberg equilibrium P value < 10^-6, minor allele counts < 20, MAF < 0.01, imputation quality < 0.8, genotype missingness > 0.1, and excluded individuals who had a genotype missing rate > 0.1.

Genome-wide Genetic Correlations Between Traits

We estimated the heritability resulting from genotyped single-nucleotide polymorphisms (SNPs) and genetic correlations between traits using LD Score Regression based on GWAS summary statistics for the European ancestry population (24, 25). Because of the smaller sample size, we used bivariate GREML analysis from GCTA, which relies a genetic relatedness matrix rather than a reference population, to estimate the heritability and genetic correlation between traits among African ancestry women (23).

Meta-analysis

We performed fixed effects inverse-variance weighted meta-analysis to conduct cross-ancestry GWAS of each sex hormone using the METAL software (26). We used the Cochran Q test as implemented in METAL to test for heterogeneity of effects between European and African ancestries. We present the ratio of beta coefficients between African ancestry and European ancestry women as a function of their meta-analysis P values (9, 15). We additionally conducted genome-wide SNP-specific tests for heterogeneity to identify genetic variants with statistically significant differences in effect size by menopausal status in the European-ancestry GWAS based on the results of the Cochran Q-test as implemented in METAL (26). For all genome-wide analyses, we considered P < 5 × 10^-8 as statistically significant.

Gene-level Association Analysis and Tissue-specific Enrichment

Gene-level association analysis was performed using the multimarker analysis of genomic annotation (MAGMA) (27) tool implemented in the Functional Mapping program for GWAS results. MAGMA analyzes SNP-level summary genetic associations while taking linkage disequilibrium (LD) between SNPs into account (28). SNPs were mapped to a gene if they were located between the start and end sites of the gene. All analyses were based on the human genome build 37.

MAGMA “gene-property analysis” was performed to test for associations between gene-level signals and tissue-specific gene expression profiles. Gene-property analysis uses a multivariable regression model that includes gene expression in a specific tissue type and the average gene expression across 54 tissue types to evaluate the relationship between tissue specificity and gene-level association. Tissue-specific gene expression data were from GTEx version 7 (29). We performed gene property analysis for tissue specificity of testosterone and SHBG using the GWAS results from women of European ancestry overall and by menopausal status.

Results

Our final sample sizes were 229 966 (98%) European ancestry women and 5336 (2%) African ancestry women. Among those empirically assigned as European ancestry, 51 081 were premenopausal and 84 194 postmenopausal. Of those women empirically assigned as African ancestry, 1994 were premenopausal and 1200 were postmenopausal. Because of the relatively small sample sizes, results within menopausal status for African ancestry women are only reported in the supplementary data (30).

As previously mentioned, we observed a greater proportion of premenopausal women with detectable levels of estradiol compared with postmenopausal women (Table 1). We also observed greater mean concentrations of testosterone and SHBG in premenopausal women compared with postmenopausal women, and across both pre- and postmenopausal women, we observed right-skewed distributions of sex hormones as indicated by the lower median compared with the mean and in relation to the interquartile range (Table 1). Although women of African ancestry had lower mean concentrations of all 3 sex hormones, differences were not statistically significant.

Table 1.

Descriptive characteristics of sex hormone concentrations in pre- and postmenopausal women in the UK Biobank population

Premenopausal (N = 62 587)		Postmenopausal (N = 124 820)
Mean (SD)	Median (IQR)	Mean (SD)	Median (IQR)
Estradiol: n = 39 586 (63.2%)		Estradiol: n = 4,262 (3.4%)
571 (448)	432 (290, 683)	339 (280)	257 (204, 376)
Testosterone: n = 53 179 (85.0%)		Testosterone: n = 95 031 (76.1%)
1.22 (0.60)	1.13 (0.82, 1.49)	1.08 (0.67)	0.97 (0.69, 1.32)
SHBG: n = 52 127 (83.3%)		SHBG: n = 105 289 (84.4%)
68.0 (33.1)	62.5 (44.2, 84.9)	59.5 (28.1)	54.8 (39.4, 74.2)

n represents the number available for inclusion in GWAS of hormone concentrations as a continuous trait. Estradiol as a dichotomized trait indicating detectable measures of estradiol was used as an additional phenotype for GWAS.

Abbreviations: GWAS, genome-wide association study; IQR, interquartile range.

Estradiol

We analyzed estradiol as a continuous phenotype (pmol/L) both overall and stratified by menopausal status. Among women with detectable concentrations, 33 032 European ancestry women were classified as premenopausal and 3,759 as postmenopausal (Supplementary Table 1). We identified 1 locus (lead SNP rs727428) in the 17p13.1 region near the SHBG gene that was significant in both the overall (P = 2.1 × 10^-10) analysis and when restricted to premenopausal women only (P = 1.04 × 10^-8) (Table 2; Supplementary Fig. 1a-b (30)). Among African ancestry women, 1706 had detectable concentrations of estradiol, of which 1285 were classified as premenopausal and only 76 were included in postmenopausal specific GWAS. No variants reached genome-wide significance in the African ancestry GWAS. The lead SNP, rs727428, from the European ancestry analysis only showed moderate association (P = 8 × 10^-3) in the African ancestry analysis (Supplementary Fig. 1c-d (30)). We observed no evidence of heterogeneity in effect estimates between European and African ancestry women among overlapping variants (Supplementary Table 1 (30)).

Table 2.

Independent signals from GWASs of estradiol in women of European ancestry

Phenotype	Lead variant	MAF	Region	Nearest gene	Overall		Premenopausal		Postmenopausal
					Effect estimate (SE)	P	Effect estimate (SE)	P	Effect estimate (SE)	P
Estradiol continuous, pmol/L	rs727428	0.45	17p13.1	SHBG	-0.04 (0.006)	2.1E-10	-0.05 (0.008)	1.04E-8	-0.04 (0.02)	0.08
Detectable estradiol, binary	rs2454949	0.49	5q35.2	ZNF346	0.02 (0.003)	5.9E-15	0.04 (0.006)	3.3E-9	0.03 (0.005)	2.4E-8
	rs75770066	0.03	12q14.3	HELB	0.06 (0.008)	5.4E-14	0.11 (0.02)	6.9E-11	0.10 (0.01)	9.1E-13
	rs71181755	0.47	19q13.42	TMEM150B	-0.03 (0.003)	7.5E-19	-0.04 (0.006)	2.9E-11	-0.03 (0.005)	8.2E-12
	rs16991615	0.06	20p12.3	MCM8	0.06 (0.006)	3.9E-25	0.10 (0.01)	1.8E-17	0.10 (0.01)	4.1E-24

Abbreviations: GWAS, genome-wide association study; MAF, minor allele frequency; SE, standard error.

Using estradiol as a dichotomized phenotype for detectable concentrations (yes/no) in 229 966 women of European ancestry, we report 10 loci that reached genome-wide significance in the overall analysis (Supplementary Table 3 (30)). Among these 10 loci, 2 were significant in postmenopausal but not premenopausal women (4q21.23, 8p11.23) and 4 were significant in both pre- and postmenopausal women (5q35.2, 12q14.3, 19q13.42, 20p12.3)(Table 2; Fig. 1; Supplementary Fig. 2 (30)). No loci were found to be associated with detectable concentrations of estradiol in premenopausal women but not in postmenopausal women.

Figure 1.

Genome-wide analysis of detectable concentrations of estradiol in women of European ancestry overall, and by menopausal status. Horizontal black line indicates Bonferroni correction threshold at P value < 5E-8. Points in green indicate variants with P value below a suggestive threshold of 1E-6; red points indicate variants below Bonferroni correction P value.

We included 5336 women in the African-ancestry GWAS of detectable concentrations of estradiol. Of the 10 SNPs identified in European ancestry, only 4 were included in our African ancestry analyses after quality control (Supplementary Table 2 (30); Supplementary Fig. 2b-c) (30). Of those 4 SNPs, 3 were directionally consistent with the overall European ancestry GWAS but none had P values < 0.05 or showed evidence of heterogeneity of effect across ancestries (Supplementary Table 2 (30)).

Testosterone

We conducted GWAS of circulating testosterone concentrations in 182 648 women of European ancestry, of whom 44 401 were premenopausal and 84 003 were postmenopausal. Of the 226 variants reported by Ruth et al for testosterone in women and observed in our analyses after quality control, 103 variants reached statistical significance in either pre- or postmenopausal only analyses. Of those 103 variants, 14 were significant in both premenopausal- and postmenopausal-specific analyses, 9 were significant in premenopausal but not postmenopausal women, and 72 were significant in postmenopausal but not premenopausal women (Supplementary Table 3; Supplementary Fig. 3a-b (30)).

We next tested for heterogeneity of SNP effects on testosterone between pre- and postmenopausal women of European ancestry. We identified a genome-wide significant signal for rs45446698 in the 7q22.1 locus near the CYP3A7 gene, which is involved in cholesterol and steroid metabolism (Fig. 2; Fig. 3; Supplementary Fig. 3c-d (30)). There was nearly a 2-fold increase in the effect estimate of this variant among premenopausal women compared with postmenopausal women (Table 3).

Figure 2.

Genome-wide analysis of heterogeneity of effect for testosterone in women of European ancestry between premenopausal and postmenopausal. Horizontal black line indicates Bonferroni correction threshold at P value < 5E-8. Points in green indicate variants with P value below a suggestive threshold of 1E-6; red points indicate variants below Bonferroni correction P value.

Figure 3.

Regional association plot of the chromosome 7 region (CYP3A7) showing heterogeneous effects in pre- and postmenopausal concentrations of testosterone.

Table 3.

Independent significant signals from GWAS of heterogeneity of effect between pre- and postmenopausal women of European ancestry for concentrations of testosterone and SHBG

Phenotype	Lead variant	MAF	Loci	Nearest gene	Premenopausal		Postmenopausal
					Effect estimate (SE)	P	Effect estimate (SE)	P
Testosterone, ng/dL	rs45446698	0.04	7q22.1	CYP3A7	0.44 (0.02)	6.3E-133	0.28 (0.01)	1.2E-104
SHBG, ng/dL	rs3812617	0.16	10p15.1	AKR1C4	-0.10 (0.009)	4.3E-26	-0.02 (0.006)	2.4E-4

Abbreviations: GWAS, genome-wide association study; MAF, minor allele frequency; SE, standard error.

We observe strong evidence for adrenal gland tissue specificity for testosterone in premenopausal women (β = 0.04, standard error [SE] = 0.01, P < 3 × 10^-5), but weaker evidence in postmenopausal women (β = 0.008, SE = 0.01, P = 7.8 × 10^-3) (Fig. 4; Supplementary Table 8 (30)). We conversely observed strong evidence for adipose tissue specificity in postmenopausal women (β = 0.03, SE =.01, P = 0.03) but no evidence in premenopausal women β = -6 × 10³, SE = 0.01, P = 0.67).

Figure 4.

Gene property analysis for tissue specificity of testosterone and SHBG in women of European ancestry by menopausal status. *Statistical significance at P < 0.05.

We had GWAS and testosterone data for 4229 African ancestry women. We observed no variants that reached genome-wide significance (Supplementary Fig. 3e-f (30)). We had data for 162/226 previously identified variants in European ancestry women (9), and of those 96 (59%) had directionally consistent effect estimates with women of European ancestry and 14 had P < 0.05, including rs45446698, which tags the CYP3A7 gene (Fig. 5).

Figure 5.

African and European cross-ancestry consistency of direction and ratio of size of effect for previously identified variants of circulating testosterone in women overall. Points below the x axis indicate opposite directions of effect between African and European ancestry populations. The dashed horizontal line indicates the expected ratio of equal effect sizes between ancestry populations. X axis P value is the meta-analyzed P value using genome-wide association studies from European and African ancestries in women overall. Note: The x axis was segmented because the smallest P value in the overall ancestry of women of European ancestry obscured the observations for variants with significant but relatively larger P values.

SHBG

We conducted GWAS of circulating SHBG concentrations in 196 901 women of European ancestry, of whom 43 477 were premenopausal, and 92,911 were postmenopausal. Of the 321 variants reported by Ruth et al to be associated with SHBG in women and observed in our analyses, 29 were genome-wide significant in both premenopausal specific and postmenopausal specific analyses, 2 were genome-wide significant in premenopausal but not postmenopausal women, and 16 were genome-wide significant in postmenopausal but not premenopausal women (Supplementary Table 4 (30); Supplementary Fig. 4a-b (30)).

We tested for heterogeneity of SNP effects by menopausal status and identified a signal at 10q15.1 near AKR1C4 (Fig. 6; Fig. 7; Supplementary Fig. 4C-D (30)) for which there was a statistically significant effect in premenopausal women (β = -0.10, SE =.009; P = 4.3 × 10^-26) but no association in postmenopausal women (β = -0.02, SE = 0.006; P = 2.4 × 10^-4) (Table 3). The effect estimate was approximately 5-fold higher in premenopausal women compared with the effect estimate in postmenopausal women (Table 3).

Figure 6.

Genome-wide analysis of heterogeneity of effect for SHBG in women of European ancestry between premenopausal and postmenopausal. Horizontal black line indicates Bonferroni correction threshold at P < 5E-8. Points in green indicate variants with P value below a suggestive threshold of 1E-6; red points indicate variants below Bonferroni correction P value.

Figure 7.

Regional association plot of the chromosome 10 region showing heterogeneous effects in pre- and postmenopausal concentrations of SHBG.

Our gene analyses for tissue specificity of SHBG showed strong evidence of liver tissue specificity across both pre- and postmenopausal women (Fig. 4). At a significance level of P < 0.05, we additionally observed enrichment in muscle tissue in premenopausal women (β = 0.01, SE = 0.008; P = 0.04) and in thyroid tissue in postmenopausal women (β = 0.03, SE = 0.01; P = 0.02).

We included 4522 women of African ancestry in our GWAS of concentrations of SHBG. Of the previously reported 321 variants, we observed 209 in the African-ancestry analyses, of which 148 (71%) were directionally consistent with the European-ancestry effect estimates. Of the 209 previously reported variants, 36 had P < 0.05 (Supplementary Table 4 (30)). In particular, we observed a genome-wide significant signal for the SHBG gene with rs727428 as the lead variant (β = 0.23, SE = 0.02; P = 7.6 × 10^-29). We present the ratio of effect estimates for the 209 variants observed in European and African ancestry populations (Fig. 8).

Figure 8.

Ratio of beta coefficients between African and European GWAS of SHBG in all women by meta-analysis P value for previous reported variants. Points below the x axis indicate opposite directions of effect between African and European ancestry populations. The dashed horizontal line indicates the expected ratio of equal effect sizes between ancestry populations. The x axis P value is the meta-analyzed P value.

Genetic Correlations Between Concentrations of Circulating Hormones

We estimated the heritability and genome-wide genetic correlations of circulating concentrations of sex hormone traits overall and by menopausal status for women of European ancestry (25, 31). Overall SNP heritability estimates for testosterone and SHBG were similar to those reported by Ruth et al, and generally increased when stratified by menopausal status (Table 4). Notably, as a binary phenotype indicating detectable concentrations, estradiol had fairly low heritability (1.5%; SE = 0.002) that increased when stratified by menopausal status to 3.9% (SE = 0.01) and 3.5% (SE = 0.006) for pre- and postmenopausal concentrations, respectively. We then estimated pair-wise genome-wide genetic correlations (ie, the correlation of SNP effect estimates on 2 phenotypes across the genome). Both SHBG and testosterone showed strong genetic correlations between pre- and postmenopausal concentrations (r_g = 0.89, SE = 0.03 for SHBG; r_g = 0.93, SE = 0.07 for testosterone), whereas the genetic correlation between pre- and postmenopausal detectable estradiol concentrations was only 0.55 (SE = 0.16), suggesting that some of the underlying genetic architecture is distinct. We did not estimate SNP heritability and genetic correlations for estradiol as a continuous trait because of the low sample size. We used bivariate GREML to estimate heritability and genetic correlation of the same traits in women of African ancestry. Although estimates for estradiol were unstable because of the small sample size with detectable levels, the overall heritability estimates for testosterone and SHBG in the African-ancestry population were similar to those of European ancestry (h² = 0.23, SE = 0.07 and h² = 0.26, SE = 0.07, respectively) (Supplementary Table 7 (30)).

Table 4.

Genetic correlation between overall and menopausal status-specific hormone concentrations among European ancestry women

Trait	H2 % (SEM %)	E2	Pre	Post	E	Pre	T	Pre	Post	SHBG	Pre	Post
E2	1.5 (0.2)	-	0.91 (0.08)	0.48 (0.10)	0.51 (0.32)	0.03 (0.26)	0.09 (0.05)	0.18 (0.08)	-0.003 (0.07)	0.19 (0.05)	0.2 (0.07)	0.14 (0.06)
Premenopausal	3.9 (1.0)		-	0.55 (0.16)	0.09 (0.32)	-0.34 (0.34)	0.05 (0.07)	-0.03 (0.11)	0.006 (0.09)	0.18 (0.07)	0.21 (0.08)	0.16 (0.08)
Postmenopausal	3.5 (0.6)			-	0.45 (0.35)	-0.08 (0.26)	0.21 (0.05)	0.31 (0.08)	0.13 (0.08)	0.08 (0.06)	0.13 (0.08)	0.07 (0.06)
E	1.1 (1.0)				-	0.86 (0.21)	-0.14 (0.14)	-0.31 (0.28)	-0.17 (0.19)	0.62 (0.31)	0.8 (0.37)	0.68 (0.33)
Premenopausal	1.6 (1.3)					-	-0.05 (0.12)	-0.06 (0.20)	-0.05 (0.17)	0.6 (0.28)	0.76 (0.32)	0.67 (0.31)
Postmenopausal	NA						-	-	-	-	-	-
T	11 (0.9)						-	0.98 (0.03)	0.99 (0.02)	0.03 (0.04)	-0.07 (0.05)	-0.03 (0.04)
Premenopausal	12 (1.8)							-	0.93 (0.07)	0.08 (0.05)	-0.03 (0.07)	-0.02 (0.06)
Postmenopausal	11 (1.1)								-	-0.002 (0.04)	-0.10 (0.05)	-0.05 (0.05)
SHBG	19 (1.9)									-	0.75 (0.03)	0.85 (0.02)
Premenopausal	23 (2.8)										-	0.89 (0.03)
Postmenopausal	22 (2.6)											-

Bold correlation estimates indicate Bonferroni statistical significance (P < 0.05/50).

Abbreviations: E, estradiol; E2, estradiol dichotomized; NA, not applicable; T, testosterone.

Discussion

In this study, we investigated the genetic architecture of circulating concentrations of sex hormones by menopausal status. We identified 2 loci showing differential effects on circulating hormone concentrations by menopausal status in women of European ancestry. The CYP3A7 locus, which is the most significant region associated with concentrations of testosterone, showed a stronger effect in premenopausal women compared with postmenopausal women (β = 0.44; SE = 0.02 vs β 0.28; SE = 0.01, respectively). Although CYP3A7 is not a novel finding, the significant difference in effect sizes between pre- and postmenopausal women provides important insight into the different biological mechanisms underlying pre- and postmenopausal testosterone concentrations. The effect of rs45446698, which tags the CYP3A7*1C allele, was almost twice as large in premenopausal women compared with postmenopausal women. This SNP has previously been associated with urinary estrone glucuronide concentrations and breast cancer mortality (32).

SNPs close to the aldo-keto reductase 1C4 (AKR1C4) gene, which is involved in progesterone catabolism, showed stronger association with premenopausal concentrations of SHBG compared with postmenopausal concentrations. Previous research conducted in clinical trials of estrogen therapy found that polymorphisms in AKR1C4 were associated with mammographic density change in the combined estrogen and progestin therapy group but not in the placebo or estrogen-only group (33). Thus, effects of AKR1C4 may be dependent on higher concentrations of progesterone or progesterone derivatives; thus, we observed little to no effect in postmenopausal women because we excluded anyone with a history of menopausal hormone therapy use.

We found that genes expressed in adipose tissue were enriched in our MAGMA analysis of circulating testosterone in postmenopausal women but not in premenopausal women, where instead adrenal gland tissue were enriched. This suggests that with the loss of ovarian function following menopause, testosterone may be more regulated by adipose tissue. In contrast, we observed consistent enrichment of liver tissue in our SHBG analyses for both pre- and postmenopausal women.

We did not observe any genome-wide significant associations for concentrations of estradiol or testosterone in women of African ancestry, and this is likely because of our limited sample size (N = 5336). To contextualize this limitation, with 5000 samples, we would have a power of 0.33 to detect a variant that explains 0.5% of the variance of a trait. The relatively small sample size underscores the need for more equitable data collection of people across ancestries. There was a statistically significant signal for SHBG for known variants based on previous studies among women of European ancestry in the SHBG gene (34). We did not find evidence of any heterogeneity of effects between European ancestry and African ancestry women for SNPs previously associated with concentrations of testosterone and SHBG based on tests of heterogeneity of effects. We compared the effect estimates of SNPs identified in European ancestry women with the estimates in African ancestry women and observed a smaller difference in effect estimates with stronger discovery P value, suggesting that the strongest genetic determinants are likely the most generalizable across ancestries. Unfortunately, approximately 20% (54/229) of SNPs previously associated with testosterone concentrations were excluded in the quality controls because of high missing rates following standard quality control thresholds (35).

The previous work by Ruth et al discussed the limitations of analyzing estradiol for women in the UK Biobank because of the relatively low threshold for detection, which resulted in a substantial proportion of women missing measures for estradiol, disproportionately affecting postmenopausal women (9). We address the bias for detecting signals associated with menopausal status rather than estradiol by looking at pre- and postmenopausal women separately. In our analysis of postmenopausal women, we adjusted our analyses for time since menopause and excluded women who reported ever having taken menopausal hormone therapy. We modeled estradiol in 2 ways. We first restricted our analyses to women with detectable concentrations only and treated estradiol as a continuous outcome. We observed a genome-wide significant association close to the SHBG gene, which was associated with both overall and premenopausal concentrations. The lead SNP rs727428 is located downstream of the SHBG gene and has previously been associated with concentrations of SHBG but not with estradiol (36). Second, we modeled estradiol as a binary variable indicating detectable vs undetectable concentrations (9). Using this binary phenotype, we focused our results on signals in premenopausal women, limiting the potential bias introduced by differences in menopausal status. Four loci were associated with having detectable concentrations of estradiol in the overall and both pre- and postmenopausal only analyses. Three of these loci are near genes (12q14.3, HELB; 19q13.42, TMEM150B; and 20p12.3, MCM8) that have previously associated with age at natural menopause, primary ovarian insufficiency, and early menopause through mechanisms related to DNA repair (37). The lead variant in the 5q35.2 locus is in strong LD with variants previously associated with hormonally driven uterine fibroids and heavy menstrual bleeding (38). Of note, we did not observe associations for variants near the CYP1B1, which has been used previously in candidate gene studies of aromatase inhibitors for hormone-driven diseases (39, 40). Sensitivity analyses adjusting for factors related to timing of menstrual cycle in premenopausal women did not identify any additional signals. Although several of these genes have been previously associated with ovarian aging, their biologic functions are unclear and merit further investigation. It is possible that a more sensitive assay would provide more precise measures of estradiol for women with low concentrations, specifically for those who are postmenopausal.

The relatively low genetic correlation between a continuous measure and a binary indicator for detectable concentrations of estradiol point to the importance of improving assay sensitivity because these are not exchangeable traits. Additionally, unlike the high genetic correlations observed for premenopausal and postmenopausal testosterone, and premenopausal and postmenopausal SHBG, which were close to 1 (ie, perfectly shared genetic architecture), we saw relatively low genetic correlation for premenopausal detectable concentrations of estradiol and postmenopausal detectable concentrations of estradiol. The low shared genetic architecture suggests that there are genetic determinants of this traits within menopausal status.

We used more stringent thresholds for genetic variant inclusion in our GWAS than in previous studies of sex hormones (eg, minor allele frequency, removing indels and chromosome X), resulting in a higher proportion of missing SNPs compared with a previous publication (9). For example, of 358 variants identified by Ruth and colleagues for SHBG in women, only 321 were included in our analyses. Additionally, we observed some discrepancies in findings between our GWAS and theirs. This could be due to different adjusting factors because Ruth and colleagues additionally adjusted for time since fasting. Previous work suggested that inclusion of fasting or dilution factors did not improve bias estimates (41).

In summary, we identified novel genetic variants associated with concentrations of estradiol and genetic variants showing differential effects on circulating testosterone and SHBG according to menopausal status. Our work provides support for stratifying analyses within relevant sample characteristics, such as menopausal status, when investigating the genetic architecture of biomarkers.

Glossary

Abbreviations

1KGP

1000 Genome Project

GWAS

genome-wide association study

LD

linkage disequilibrium

MAF

minor allele frequency

MAGMA

multimarker analysis of genomic annotation

SE

standard error

SNP

single-nucleotide polymorphism

Funding

Research for this publication was supported by National Cancer Institute of the National Institutes of Health award number T32 CA094880 (“Cancer Prevention Training: Epidemiology, Nutrition, Genetics & Survivorship”) and Grant CA244670.

Disclosures

The authors do not have any conflicts of interest to disclose.

Data Availability

Genotype data analyzed during this study are in the data repositories listed in reference (42).