Observational associations in UK Biobank

UK Biobank (UKBB) is a prospective UK-based cohort study with approximately 500,000 participants aged 40–69 years at recruitment on whom a range of medical, environmental, and genetic information is collected [26]. We included 257,193 individuals self-identifying as females of white ancestry in UKBB in our analyses. Baseline measurements of BMI (total body weight [kg]/standing height squared [m²]) and waist-to-hip ratio (WHR) (waist circumference [WC] [cm]/hip circumference [HC] [cm]), and WHR adjusted for BMI (WHRadjBMI) were used to estimate general obesity (BMI) and central obesity (WHR and WHRadjBMI). In response to peer review comments, comparative body size at age 10 years (as self-reported in a questionnaire with the options ‘thinner’, ‘plumper’, or ‘about average’) was used to estimate adiposity at an earlier time point, i.e., before diagnosis of most reproductive disorders. Cases of reproductive conditions were identified based on ICD-9 and ICD-10 primary and secondary diagnoses from hospital inpatient data, self-reported illness codes, and primary care records (Table 1). We fitted logistic regression models to estimate the associations of BMI, WHR, and WHRadjBMI with prevalence of endometriosis (7,703 cases, 249,490 controls); heavy menstrual bleeding (HMB) (17,229 cases, 239,964 controls); infertility (2,194 cases, 254,999 controls); self-reported stillbirth, spontaneous miscarriage, or termination (81,102 cases, 176,091 controls); PCOS (746 cases, 256,447 controls); pre-eclampsia (2,242 cases, 254,951 controls); and UF (19,192 cases, 238,001 controls). Case definitions for pre-eclampsia included eclampsia cases to capture cases in which the former may have developed into the latter. For each disease, individuals not included in the case group were used as controls. BMI, WHR, and WHRadjBMI were adjusted for age, age squared, assessment centre, and smoking status. The residuals were rank-based inverse normally transformed. Multiple testing correction for 21 tests (3 exposures × 7 outcomes) was applied using the false discovery rate (FDR) to evaluate statistical significance while minimising false negatives [27,28].

Download:

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

Table 1. Summary of female reproductive disorders in UK Biobank.

https://doi.org/10.1371/journal.pmed.1003679.t001

We also tested associations without adjustment for smoking status, as it has previously been suggested that higher BMI increases risk of smoking [29] and adjustment for both could therefore induce collider bias. Adjustment for menopause status was not performed as up to 42% of women with reproductive disorders in UKBB report being unsure of their menopause status, as compared to 16% of women who do not have a recorded history or presence of a reproductive condition (Table 1).

To evaluate the presence of non-linear observational associations between obesity and each reproductive trait, fractional polynomial regression following the closed test procedure was performed using the mfp v1.5.2 R package [30]. This algorithm tests for the presence of an overall association, determines the likelihood of non-linearity, and selects the best-fitting fractional polynomial function. We also fitted GAMs to the same data, allowing for smoothing of the obesity trait with splines, using the mgcv 1.8–31 R package [31]. These models allow a greater degree of flexibility in modelling curves that cannot be represented by polynomials of the nth degree; however, they are also more complex and thus less immediately interpretable [32]. All models were adjusted for age, age squared, assessment centre, and smoking status. Model fits were compared with Akaike’s information criterion (AIC), with lower AIC indicating better fit [33].

As a data-driven investigation, the analyses upon which this paper is based were planned and designed for each of the sections independently; no overall prospective analysis plan was followed. Analyses dependent upon results from other sections, such as the mediation analysis dependent on associations from 2-sample MR, are noted as such within each section. This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology Using Mendelian Randomization (STROBE-MR) guideline (S1 Checklist).

Two-sample MR

Genetic instruments for BMI, WHR, and WHRadjBMI were selected based on the sentinel variants at genome-wide significant loci (P < 5 × 10⁻⁹ to account for denser imputation data) reported in the largest publicly available European ancestry genome-wide association studies (GWASs), which are meta-analyses of the Genetic Investigation of ANthropometric Traits (GIANT) and UKBB (maximum N = 806,801 individuals) [34]. Similarly, genetic instruments for predicted visceral adipose tissue (VAT) mass (N = 325,153 individuals) [35], WC (N = 462,166 individuals), HC (N = 462,117 individuals) [36], and waist-specific and hip-specific WHR (i.e., WHR instruments with specific effects on WC but not HC and vice versa; N = 18,330 individuals) [37] were selected based on the largest publicly available GWASs.

Three instrument weighting strategies were considered where sex-stratified GWAS results were available: (i) SNPs from combined-sex GWASs with combined-sex weights (effect sizes), (ii) combined-sex SNPs with female-specific weights, and (iii) female-specific SNPs with female-specific weights. The method of female-specific SNPs with female-specific weights produced the strongest instruments as evaluated by F-statistics, i.e., mean β²/σ² over all SNPs in the instrument, and was thus chosen for analysis (S1 Table). Additionally, due to concerns of ascertainment bias in UKBB [38,39], sensitivity analyses with combined-sex instruments (combined-sex SNPs with combined-sex weights) were also performed. In response to peer review comments, an additional sensitivity analysis with SNPs from previous GIANT releases for BMI [40] and WHR and WHRadjBMI [41] that do not include UKBB participants was also performed to alleviate potential ascertainment bias. Finally, to alleviate concerns of collider bias in the WHRadjBMI GWAS [42], we constructed a joint WHR and BMI instrument to perform multivariable MR.

Associations of the genetic instruments for obesity traits with female reproductive diseases were obtained by performing a fixed-effect inverse-variance-weighted (IVW) meta-analysis of publicly available GWAS summary statistics from 2 large biobank projects—FinnGen and UKBB [43]. The meta-analysis was performed using METAL [44] by matching the relevant ICD codes (S2 Table) for the following traits: infertility (4,996 cases, 421,223 controls), pre-eclampsia (2,711 cases, 480,373 controls), and UF (21,835 cases, 456,551 controls). For endometriosis, summary statistics were obtained by request from a recent European ancestry GWAS [45] and meta-analysed as above with publicly available FinnGen and UKBB summary statistics (12,210 cases, 450,183 controls). For HMB (9,813 cases, 210,946 controls), sporadic miscarriage (i.e., 1–2 miscarriages; 50,060 cases, 174,109 controls), and multiple consecutive miscarriage (i.e., ≥3 consecutive miscarriages; 750 cases, 150,215 controls), publicly available summary statistics were obtained from recent European ancestry GWASs that include UKBB individuals [3,46]. For PCOS, estimates were based on a fixed-effect IVW meta-analysis of published GWAS summary statistics [47], publicly available GWAS results by FinnGen, and a European ancestry GWAS run in UKBB using SAIGE (11,186 cases, 273,812 controls) (S3 Table). As a sensitivity analysis, all MR tests were performed using disease association estimates based on FinnGen only, where available, to alleviate bias due to sample overlap between the exposure and outcome GWAS sources.

Power to detect MR associations was calculated using 2 methods, one designed for general 2-sample MR [48] and the other for MR performed on binary outcomes [49]. Briefly, these methods calculate power by accounting for GWAS sample size, the proportion of cases in case–control GWASs, and the variance explained by genetic instruments for the exposure. The power to detect a true odds ratio (OR) association of 1.1 or more extreme at an unadjusted significance level of 0.05 was estimated.

Instrument SNPs were extracted from the outcome GWAS results and harmonised for consistency in the alleles, and MR was performed using the TwoSample MR v0.5.4 R package [50]. Two methods for MR—IVW and MR-Egger—were evaluated, and the best method was selected via Ruecker’s framework [51]. Briefly, this framework advises to choose the MR method with least heterogeneity as assessed by Cochran’s Q statistic, while accounting for the trade-off between power and pleiotropy [52]. IVW results, which were chosen by Ruecker’s framework for all tested associations, are reported in the main text. However, as these are still susceptible to pleiotropy, results from all methods, including weighted median MR [53], are calculated for robustness and displayed in the supplementary information. Correction for multiple hypothesis testing for 24 tests in each analysis (3 exposures × 8 outcomes) was applied with the FDR method, and significance established at FDR < 0.05. MR-Egger intercept tests were performed to detect horizontal pleiotropy, and single-SNP and leave-one-out analyses were used to identify outlier SNPs driving relationships [50].

Reverse MR for obesity traits regressed on female reproductive conditions was performed as detailed above. Genetic instruments for endometriosis (14,926 cases, 189,715 controls) [54], PCOS (10,174 cases, 103,164 controls) [47], and UF (20,406 cases, 223,918 controls) [3] were constructed from index variants identified by the largest European ancestry GWAS for each trait. Instrument strength was assessed by F-statistics (endometriosis, 16 SNPs, F = 5.13; PCOS, 14 SNPs, F = 41.6; UF, 29 SNPs, F = 11.1). Associations of genetic instruments for these reproductive conditions with BMI, WHR, and WHRadjBMI were obtained from female-specific summary statistics from the above-mentioned GIANT–UKBB meta-analysis [34].

Non-linear MR

For non-linear MR analyses, we selected female UKBB participants of white British ancestry with no second-degree or closer relatives in the study, as identified by the UKBB team [55], to avoid violation of the MR assumption of random assignment of genetic variants; 207,705 women were retained following this selection. Genetic instruments for BMI were constructed for each individual using female-specific index variants from Pulit et al.’s GIANT–UKBB meta-analysis [34]. The instruments for BMI explained 4.15% of trait variance after adjustment for age, age squared, smoking status, assessment centre, genotyping array, and the first 10 genetic principal components to account for population stratification. Binomial non-linear MR, a method designed to assess genetically predicted causal relationships in different exposure strata while avoiding collider bias, was performed using the fractional polynomial method with 100 quantiles and the piecewise linear method with 10 quantiles [56]. Outcomes were restricted to female reproductive disorders with prevalence > 5% in UKBB (i.e., HMB, miscarriage, and UF) to maintain sufficient sample sizes in each quantile to estimate localised average causal estimates. We assessed non-linearity with the fractional polynomial non-linearity and Cochran’s Q tests, and tested heterogeneity of the instrumental variable with the Cochran’s Q and trend tests. All analyses were performed with the nlmr v2.0 R package [56].

MR with mediation analysis

To investigate the extent to which obesity affects female reproductive disorders via hormone-related mediators, 2-step MR by the product of coefficients method was performed using GWAS summary statistics. This method was chosen as female reproductive disease phenotypes are binary outcomes with disease prevalence < 10% in UKBB, for which 2-step MR provides the least biased estimates of mediation [57]. Summary statistics for leptin (N = 33,987) [58], fasting insulin (N = 51,750) [59], and insulin sensitivity (N = 16,753) [60] were obtained from publicly available European ancestry GWAS sources that do not include samples from UKBB to minimise bias from sample overlap (S1 Table).

In the first step of 2-step MR, the mediators were regressed on obesity-related exposures using summary statistics MR methods described above. The direction of causality for all relationships was confirmed with the MR-Steiger directionality test [61], and reciprocal MR with mediator instruments and obesity-related exposures as outcomes were performed to ensure correct direction of causality. In the second step, multivariable MR was performed using combined genetic instruments for each obesity trait and hormone to estimate the independent effect of the mediator on each outcome after adjusting for the value of the exposure, and to estimate the independent effect of the exposure on outcome when adjusted for the value of each mediator. This was only done for traits where the total unadjusted effect of the exposure on the outcome was significant (FDR < 0.05). ORs for binary outcomes were converted to log ORs to calculate mediated effect by the product of coefficients method. The proportion of effect mediated was calculated by dividing the indirect effect by the total effect. Standard errors were estimated with the delta method [62].

Disease and SNP clustering

To assess similarities in the aetiological relationships of different reproductive conditions with obesity traits, we projected single SNP genetic effect estimates for BMI, WHR, and WHRadjBMI on the reproductive traits, estimated using the Wald ratio, in a 2-dimensional space using the Uniform Manifold Approximation and Projection (UMAP). Briefly, each 8-by-M matrix of SNP effect estimates (for 8 female reproductive outcomes, where M is 281 for BMI, 203 for WHR, and 266 for WHRadjBMI) was reduced to a 2-dimensional representation while maintaining as close a topological relationship between the 8 reproductive outcomes as possible, as measured by cross entropy [63]. SNPs were annotated to their nearest gene with SNPsnap [64].

To identify the genetic instruments driving the obesity–reproductive trait association, and identify clusters of SNPs with distinct associations, we clustered SNPs by the magnitude of their causal estimates using mixture model clustering in the MR-Clust v0.1.0 R package [65]. For each obesity trait–reproductive disease pair, the algorithm distinguishes the genetic instruments for the obesity traits that do not have an effect on the disease (‘null clusters’) from those that have a similar scaled effect on the disease (‘substantial clusters’) and those that have a scaled effect that cannot be grouped with other variants (‘junk clusters’).

Research ethics

This research has been conducted using the UKBB resource under application number 11867. All procedures and data collection in UKBB were approved by the UKBB Research Ethics Committee (reference number 11/NW/0274), with participants providing full written informed consent for participation in UKBB and subsequent use of their data for approved applications. All other publicly available de-identified summary data used in this study have ethical permissions from their respective institutional review boards.

Code availability

All scripts used in analyses are deposited at https://github.com/lindgrengroup/obesity_femrepr_MR.

Obesity traits are observationally associated with female reproductive diseases in UKBB

BMI at baseline assessment (age 40–69 years) was positively associated with the prevalence of most female reproductive disorders in UKBB, with the strongest association observed between BMI and PCOS (OR [95% CI] per 1 SD higher BMI = 1.87 [1.80–1.94], P = 1.90 × 10⁻⁶⁴). Associations with WHRadjBMI were null or lower than those for WHR (ORs for WHRadjBMI versus WHR: PCOS, 1.06 versus 1.48; pre-eclampsia, 1.02 versus 1.13; endometriosis, 1.02 versus 1.08; HMB, 1.06 versus 1.14; UF, 1.02 versus 1.08), indicating that BMI may be driving many of the associations between WHR and female reproductive diseases (Figs 1 and 2; Table 2). Infertility was the only disorder for which BMI (OR = 0.894, P = 2.16 × 10⁻⁰⁷) and WHR (OR [95% CI] = 0.927 [0.884–0.969], P = 4.08 × 10⁻⁰⁴) were inversely associated with disease. All associations were estimated in 257,193 women of European ancestry in UKBB, with cases defined as in Table 1 and all non-cases as controls.

Download:

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

Fig 1. Predicted probability of developing female reproductive disorders as a function of obesity-related traits.

A series of logistic regression, fractional polynomial, and generalised additive models were fitted to estimate the probability of developing female reproductive disorders as a function of obesity-related traits in UK Biobank. Predicted fits for logistic and best-fitting non-linear models were better than those for logistic regression (as evaluated with Akaike information criterion), and 95% confidence intervals about the mean are displayed. Asterisks indicate that non-linear models fit the data better than linear models. All models were adjusted for age, age squared, assessment centre, and smoking status. BMI, body mass index; spont. miscarr., spontaneous miscarriage; WHRadjBMI, waist-to-hip ratio adjusted for BMI.

https://doi.org/10.1371/journal.pmed.1003679.g001

Download:

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

Fig 2. Comparison of observational and genetically causal relationships between obesity-related traits and female reproductive disorders.

Odds ratios and 95% confidence intervals per 1 SD higher obesity trait displayed. Significant relationships (false discovery rate [FDR]–adjusted P value < 0.05) are in solid lines while non-significant (n.s.) ones are shown with dotted lines. For observational results, BMI, WHR, and WHRadjBMI adjusted for age, age squared, region (assessment centre), and smoking status are used as predictors in a logistic regression model. Causal relationships between genetically predicted obesity-related traits and female reproductive disorders are assessed by 2-sample Mendelian randomisation (MR). The displayed method (inverse-variance-weighted) was determined via Rucker’s model selection framework to minimise heterogeneity of the estimate. “Miscarriage (sporadic)” is self-reported stillbirth, spontaneous miscarriage, or termination for observational results and sporadic miscarriage for MR results. BMI, body mass index; HMB, heavy menstrual bleeding; Misc. (mult.), multiple consecutive miscarriage; PCOS, polycystic ovary syndrome; WHR, waist-to-hip ratio; WHRadjBMI, waist-to-hip ratio adjusted for BMI.

https://doi.org/10.1371/journal.pmed.1003679.g002

Download:

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

Table 2. Observational and genetic associations between obesity traits and female reproductive disorders.

https://doi.org/10.1371/journal.pmed.1003679.t002

Non-linear models explained the associations of BMI with many reproductive disorders better than linear models. We observed inverted-U and plateau relationships with endometriosis (linear AIC = 67,091, GAM AIC = 67,051), UF (linear AIC = 134,160, GAM AIC = 134,094), HMB (linear AIC = 116,687, GAM AIC = 116,636), miscarriage (linear AIC = 314,828, GAM AIC = 314,819), and pre-eclampsia (linear AIC = 24,826, GAM AIC = 24,814) (Fig 1; S4 Table). All 3 obesity traits displayed U-shaped relationships with PCOS.

We additionally found that having comparatively larger body size than average at age 10 years, as self-reported, was associated with increased prevalence of endometriosis (OR [95% CI] = 1.12 [1.06–1.19], P = 2.67 × 10⁻⁰⁴), heavy menstrual bleeding (OR [95% CI] = 1.18 [1.14–1.22], P = 3.43 × 10⁻¹⁵), PCOS (OR [95% CI] = 1.92 [1.75–2.10], P = 1.59 × 10⁻¹³), and UF (OR [95% CI] = 1.07 [1.03–1.11], P = 1.03 × 10⁻⁰³) (S6 Table). However, being thinner than average in early life was also associated with increased prevalence of endometriosis (OR [95% CI] = 1.21 [1.16–1.26], P = 1.20 × 10⁻¹³), heavy menstrual bleeding (OR [95% CI] = 1.10 [1.07–1.14], P = 7.03 × 10⁻⁰⁸), and miscarriage (OR [95% CI] = 1.09 [1.07–1.11], P = 4.03 × 10⁻¹⁸). Observational estimates for the relationship between all obesity traits and female reproductive disorders did not differ with or without adjustment for smoking status (S5 Table). Statistical significance after multiple testing correction for 21 tests (3 exposures × 7 outcomes) was established at FDR < 0.05, unadjusted P < 0.04.

Body fat distribution is genetically causally related to risk of female reproductive diseases

Two-sample MR indicated that higher genetically predicted WHR and WHRadjBMI are associated with higher risk of pre-eclampsia (OR [95% CI] per 1-SD increase in trait: WHR, 1.57 [1.16–2.10], P = 2.92 × 10⁻⁰³; WHRadjBMI, 1.43 [1.13–1.80], P = 2.46 × 10⁻⁰³), HMB (WHR, 1.28 [1.13–1.46], P = 1.42 × 10⁻⁰⁴; WHRadjBMI, 1.16 [1.04–1.29], P = 5.20 × 10⁻⁰³), endometriosis (WHR, 1.24 [1.05–1.47], P = 1.00 × 10⁻⁰²; WHRadjBMI, 1.24 [1.08–1.41], P = 1.67 × 10⁻⁰³), UF (WHR, 1.24 [1.10–1.41], P = 6.20 × 10⁻⁰⁴; WHRadjBMI, 1.17 [1.06–1.29], P = 1.95 × 10⁻⁰³), infertility (WHRadjBMI, 1.21 [1.03–1.43], P = 2.14 × 10⁻⁰²), and PCOS (WHR, 1.07 [1.02–1.11], P = 4.30 × 10⁻⁰³) (Table 2; Fig 2). The genetic estimates for the effect of WHR and WHRadjBMI on several reproductive disorders were higher than their observational counterparts (heterogeneity P-het < 0.0151 for 6/14 associations). While genetically predicted BMI was also associated with increased risk of most female reproductive disorders (ORs per 1 SD higher BMI ranged from 1.06 for sporadic miscarriage to 2.09 for pre-eclampsia), MR estimates of associations between BMI and PCOS (OR [95% CI] = 1.13 [1.08–1.19], P = 7.60 × 10⁻⁰⁸) were much attenuated compared to observational results (P-het = 2.82 × 10⁻³⁰).

Genetically predicted VAT mass was associated with the development of pre-eclampsia (OR [95% CI] per 1-kg increase in predicted VAT mass = 3.08 [1.98–4.79], P = 6.65 × 10⁻⁰⁷), PCOS (OR [95% CI] = 1.15 [1.08–1.23], P = 3.24 × 10⁻⁰⁵), and HMB (OR [95% CI] = 1.32 [1.06–1.64], P = 0.0125) (Fig 3; S7 Table). The differential association of genetically predicted body fat distribution with female reproductive traits was further reflected in the heterogeneous associations of WC and HC with disease development. Increased WC posed a higher risk than did increased HC for pre-eclampsia (OR per 1-SD increase: 1.93 for WC versus 1.40 for HC, heterogeneity P-het = 0.0373), HMB (1.41 for WC versus 1.12 for HC, P-het = 3.60 × 10⁻⁰³), UF (1.32 for WC versus 1.12 for HC, P-het = 7.70 × 10⁻⁰³), and PCOS (1.16 for WC versus 1.10 for HC, P-het = 0.0325). We did not see this heterogeneity in observational associations (all P-het > 0.164) (Fig 3; S12 Table).

Download:

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

Fig 3. Causal associations of genetically predicted visceral adipose tissue mass, waist circumference, and hip circumference with female reproductive disorders.

Odds ratios (ORs) per 1-kg increase in predicted visceral adipose tissue (VAT) mass or per 1-SD increase in waist circumference or hip circumference and 95% confidence intervals (CI) are estimated by 2-sample Mendelian randomisation. The displayed method (inverse-variance-weighted) was determined via Rucker’s model selection framework. The number of cases for each disease is indicated. Significant relationships (false discovery rate [FDR]–adjusted P value < 0.05) are in solid lines while non-significant (n.s.) ones are shown with dotted lines. c., circumference; HMB, heavy menstrual bleeding; PCOS, polycystic ovary syndrome.

https://doi.org/10.1371/journal.pmed.1003679.g003

No significant associations were found when restricting MR analyses to genetic instruments with a specific effect on waist but not HC, or on hip but not WC (S13 Table; S1 Fig), but the power of these instruments to detect ORs more extreme than 1.1 was limited to 5%–20% (S14 Table). No non-linear MR models explained the genetic associations of BMI with any reproductive disorder better than linear MR models (S4 Fig). However, the power to detect non-linear effects was severely limited by the lower number of cases in each quantile of the BMI distribution in which analyses were run. Statistical significance after multiple testing correction for 24 tests (3 exposures × 8 outcomes) was established at FDR < 0.05, unadjusted P < 0.03.

SNPs identified in female-only GWASs and with female-specific weights for BMI, WHR, and WHRadjBMI [34] were found to be the strongest instruments, with F-statistics > 60; instrument strength for WC and HC was >45 (S1 Table). We prioritised MR results based on the IVW method over less powered MR-Egger and weighted median methods, as there was no evidence for directional pleiotropy (MR-Egger horizontal pleiotropy P > 0.0547) and the ratio of Cochran’s Q′ (Egger) to Q (IVW) was >0.876 (S7 Table). For the 24/64 analyses for which IVW indicated a significant effect, the effect estimate of the other methods was either directionally consistent (13/24) or non-significant (11/24), but never opposite (S1 Fig). Estimates were also consistent when based only on FinnGen summary statistics (heterogeneity P > 0.163), with instruments from GIANT only (heterogeneity P > 0.0541), or with combined-sex instruments (heterogeneity P > 0.999), suggesting that the findings were not substantially biased due to sample overlap between exposure and outcome GWAS sources or ascertainment bias in UKBB [35,36] (S9 and S10 Tables; S1 and S2 Figs). Finally, results for WHRadjBMI did not appear to be affected by collider bias, as estimates did not differ when using WHRadjBMI GWAS instruments compared to a multivariable analysis for WHR SNPs and BMI SNPs in the same model (S11 Table; S3 Fig).

We did not find evidence for reverse causal associations of endometriosis, PCOS, or UF with BMI, WHR, and WHRadjBMI (S15 Table). However, these estimates may be biased by weak genetic instruments for endometriosis (F-statistic = 5.13) and UF (F-statistic = 11.1) and high heterogeneity for all associations (Cochran’s Q P < 4.71 × 10⁻⁰⁶). We were limited in assessing the reverse causality of other female reproductive conditions on obesity traits by the lack of index SNPs in large-scale publicly available GWAS summary statistics.

Leptin and insulin mediate the genetically predicted causal associations of obesity with female reproductive disorders

We applied a series of MR-based mediation analyses [66,67] to study the role of hormonal factors—leptin and insulin resistance—in mediating the relationships between obesity and female reproductive health (Fig 4A). The effects of BMI, WHR, and WHRadjBMI on endometriosis, PCOS, pre-eclampsia, and UF were attenuated (95% CIs of ORs all contain 1) when adjusted for leptin, fasting insulin, or insulin sensitivity as measured by the modified Stumvoll Insulin Sensitivity Index (ISI) (Fig 4B, S16 Table). Furthermore, leptin and insulin influence risk of pre-eclampsia independently of obesity. After adjustment for BMI, leptin (β ± standard error [SE] = 0.887 ± 0.232, P = 1.28 × 10⁻⁰⁴), fasting insulin (β ± SE = 1.42 ± 0.441, P = 1.29 × 10⁻⁰³), and ISI (β ± SE = −0.503 ± 0.169, P = 2.99 × 10⁻⁰³) were all associated with risk of pre-eclampsia (Fig 4C; Table 3). Similarly, fasting insulin (β ± SE = 1.27 ± 0.473, P = 7.15 × 10⁻⁰³) and ISI (β ± SE = −0.793 ± 0.183, P = 1.46 × 10⁻⁰⁵) had genetic associations with pre-eclampsia upon adjustment for WHR. Leptin, fasting insulin, and ISI did not have significant genetic associations with endometriosis, PCOS, or UF after adjustment for obesity traits, nor did they have any associations with the obesity traits themselves (S19 Table). Statistical significance after multiple testing correction for 48 tests (3 exposures × 4 outcomes × 4 adjustments) was established at FDR < 0.05, unadjusted P < 0.01.

Download:

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

Fig 4. Hormone mediation of genetically predicted causal effect of obesity on female reproductive disorders.

(A) Method outline for 2-step Mendelian randomisation (MR) to estimate hormonally mediated effects between obesity and female reproductive disorders. In step 1, the effect of exposures on mediators is estimated using instruments for the exposure alone, while in step 2 the independent effect of the mediators on outcomes is estimated using multivariable MR (MVMR) adjusted for exposures. All SNP–phenotype effect estimates come from different genome-wide association studies. (B) Estimated effects of obesity traits on female reproductive disorders adjusted for mediators. MVMR was performed with combined genetic instruments for each exposure–mediator combination, displayed here for relationships where unadjusted (unadj.) exposure–outcome effect was significant (false discovery rate [FDR] < 0.05). (C) Example of a mediated relationship, shown here for BMI effect on pre-eclampsia. Estimated effects for exposure–mediator (betas and standard errors) and for mediator–outcome (log odds ratios and standard errors) effects are shown. Repr., reproductive; BMI, body mass index; n.s., not significant; PCOS, polycystic ovary syndrome; sensitiv., sensitivity; WHR, waist-to-hip ratio; WHRadjBMI, waist-to-hip ratio adjusted for BMI.

https://doi.org/10.1371/journal.pmed.1003679.g004

Download:

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

Table 3. Mendelian randomisation estimates of female reproductive disorders regressed on metabolic hormones, adjusted for obesity traits.

https://doi.org/10.1371/journal.pmed.1003679.t003

We calculated the proportion of total obesity effect mediated by leptin, fasting insulin, and ISI for disorders where the effects of obesity traits and mediators were significant at unadjusted P < 0.05. We found that leptin (50.2% of the effect of BMI on pre-eclampsia), fasting insulin (27.7%–36.6%), and ISI (19.1%–50.1%) each mediated the total genetically predicted effects of obesity traits on female reproductive disorders (Table 4).

Download:

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

Table 4. Proportion of effect mediated for exposure–mediator–outcome relationships.

https://doi.org/10.1371/journal.pmed.1003679.t004

Other metabolic and hormone pathways may drive the aetiological relationships of obesity with female reproductive diseases

We assessed the similarities in the aetiological relationships of different reproductive conditions with obesity, by projecting the single SNP genetic effect estimates for BMI, WHR, and WHRadjBMI on the reproductive traits in a 2-dimensional space using UMAP (Fig 5A). The UMAP projections based on all obesity traits clustered female reproductive diseases into three groups; one, consisting of endometriosis, UF, infertility, and HMB, which was separated from the second (sporadic and multiple consecutive miscarriage). Group 3 consisted of PCOS and pre-eclampsia, which clustered closely in UMAP plots of the effect of WHR and WHRadjBMI variants, but were separated by BMI-associated variants. This reflects a shared genetic component of the aetiological role of general and central obesity in the 3 groups of reproductive conditions.

Download:

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

Fig 5. Clustering of the genetic effect of obesity on female reproductive disorders.

(A) Female reproductive disorders separated by the genetic effects of different obesity traits on reproductive conditions. Uniform Manifold Approximation and Projection (UMAP) was used to plot the separation of diseases based on genetic effects of obesity instruments, evaluated by 2-sample Mendelian randomisation. (B) Number of SNPs in substantial clusters for each obesity × female reproductive condition relationship. SNPs with ≥80% probability of belonging to a substantial cluster (MRClust) are displayed with their nearest gene as annotated by SNPsnap and ‘+’ or ‘−’ representing causal effect direction. BMI, body mass index; Endo, endometriosis; HMB, heavy menstrual bleeding; Infert., infertility; M. misc., multiple consecutive miscarriage; PCOS, polycystic ovary syndrome; PE, pre-eclampsia; S. misc., sporadic miscarriage; UF, uterine fibroids; WHR, waist-to-hip ratio; WHRadjBMI, waist-to-hip ratio adjusted for BMI.

https://doi.org/10.1371/journal.pmed.1003679.g005

We further examined whether different aspects of obesity play an aetiological role in different reproductive conditions. For each obesity trait–reproductive disease pair, we grouped the genetic instruments for the obesity traits by those that do not have associations with the disease (‘null clusters’), those that have a similar scaled association with the disease (‘substantial clusters’), and those that have a scaled association that cannot be grouped with other variants (‘junk clusters’) using MRClust [65]. One substantial cluster was identified for each pair of obesity traits and reproductive conditions. The only exception to this was the pair WHRadjBMI and UF, for which 2 substantial clusters were identified, one with positive genetic association and the other with negative association (S5 Fig; S17 Table). Of the 4 SNPs in the negative effect cluster, rs2277339 (missense variant in PRIM1 and upstream of HSD17B6, involved in steroid biosynthesis) is associated with primary ovarian insufficiency, early menopause, and PCOS [68,69], and rs11694173 is intronic to THADA, which is also associated with PCOS [47]. On the other hand, 4 of 10 SNPs in the positive effect cluster are associated with metabolic traits—rs12328675 and rs2459732 are associated with circulating leptin [70], rs6905288 is associated with type 2 diabetes and thyroid stimulating hormone [71,72], and rs4686696 is intronic to insulin-like growth factor IGF2BP2.

SNPs with high probability of belonging to the substantial cluster (≥80% probability) were generally unique to each obesity–disease relationship, with no more than 2 variants shared between any 2 clusters (Fig 5B). However, 6 BMI index SNPs had positive genetic estimates for both PCOS and pre-eclampsia, including rs1121980 in the adipose-associated gene FTO and rs7498665 in SH2B1, linked to insulin resistance in obesity. The BMI-associated variant rs7084454 (intronic to MLLT10) was shared by substantial clusters for PCOS, endometriosis, and UF, while rs114760566 (mapped to HMGA1, associated with type 2 diabetes and multiple lipomatosis) was shared by endometriosis and UF. We evaluated the biological effect of the top SNPs in each substantial cluster with the DEPICT algorithms for pathway enrichment and gene prioritisation. We recapitulated the known association of the GEMIN5 subnetwork with PCOS in SNPs causal for BMI–PCOS [73]. Gene prioritisation for WHR–endometriosis causal SNPs highlighted TBX15, an important mesodermal transcription factor with roles in endometrial and ovarian cancer [74,75].