Abstract
Background
Early age at menarche (AAM) has been associated with a higher risk of
carotid artery intima‐media thickness (cIMT), an indicator of
subclinical vascular disease, albeit the mechanisms underlying this
association remain elusive. A better understanding of the relationship
between AAM, modifiable cardiometabolic risk factors, and subclinical
atherosclerosis may contribute to improved primary prevention and
cardiovascular disease treatment. We aimed to investigate the putative
causal role of AAM on cIMT, and to identify and quantify the
potentially mediatory effects of cardiometabolic risk factors
underlying this relationship.
Methods and Results
We conducted linkage disequilibrium score regression analyses between
our exposure of interest, AAM, our outcome of interest, cIMT and
potential mediators of the AAM‐cIMT association to gauge cross‐trait
genetic overlap. We considered as mediators the modifiable
anthropometric risk factors body mass index (BMI), systolic blood
pressure (SBP), lipid traits (total cholesterol, triglycerides,
high‐density lipoprotein cholesterol, and low‐density lipoprotein
cholesterol), and glycemic traits (fasting glucose). We then leveraged
the paradigm of Mendelian randomization to infer causality between AAM
and cIMT, and to identify whether cardiometabolic risk factors served
as potential mediators of this effect. Our analyses showed that
genetically predicted AAM was inversely associated with cIMT, BMI, SBP,
and triglycerides, and positively associated with high‐density
lipoprotein, low‐density lipoprotein, and total cholesterol. We showed
that the effect of genetically predicted AAM on cIMT may be partially
mediated through BMI (20.1% [95% CI, 1.4% to 38.9%]) and SBP (13.5%
[95% CI, 0.5%–26.6%]). Our cluster‐specific Mendelian randomization
revealed heterogeneous causal effect estimates of age at menarche on
BMI and SBP.
Conclusions
We highlight supporting evidence for a potential causal association
between earlier AAM and cIMT, and almost one third of the effect of AAM
on cIMT may be mediated by BMI and SBP. Early intervention aimed at
lowering BMI and hypertension may be beneficial in reducing the risk of
developing subclinical atherosclerosis due to earlier age at menarche.
Keywords: age at menarche, BMI, cIMT, Mendelian randomization, SBP
Subject Categories: Genetic, Association Studies; Cardiovascular
Disease; Epidemiology
__________________________________________________________________
Nonstandard Abbreviations and Acronyms
AAM
age at menarche
cIMT
carotid intima‐media thickness
FG
fasting glucose
IVW
inverse‐variance weighted
LDSC
linkage disequilibrium score
MR
Mendelian randomization
TC
total cholesterol
Clinical Perspective.
What Is New?
* This Mendelian randomization study provides evidence that
genetically predicted earlier age at menarche is potentially
causally associated with carotid intima‐media thickness, a marker
of subclinical atherosclerosis.
* Our results suggest that the association between age at menarche
and carotid intima‐media thickness may be partially mediated by
adult body mass index and systolic blood pressure.
What Are the Clinical Implications?
* These findings further support the adverse health effects of
earlier age at menarche.
* Early intervention aimed at lowering body mass index and
hypertension may be beneficial in reducing the risk of developing
subclinical atherosclerosis due to earlier age at menarche.
Cardiovascular disease (CVD) is one of the leading causes of morbidity
and death in women worldwide.[42] ^1 Emerging evidence suggests that
sex‐specific risk factors such as age at menarche (AAM), age at
menopause, and timing and number of births also contribute to a woman's
unique CVD risk, in addition to the traditionally established CVD risk
factors including hyperlipidemia, hypertension, diabetes, and obesity.
AAM in particular is considered to be an independent risk factor for
lifetime risk of CVD events, CVD death, and overall death in women.[43]
^2 , [44]^3 Epidemiological studies have highlighted an association
between earlier menarche (<10 years of age) in women and a
significantly higher relative risk of coronary heart disease than women
who experienced menarche at a later onset (12–13 years of age).[45] ^2
, [46]^4 These findings suggest that sex‐specific factors play an
important role in lifetime CVD risk that is potentially independent of
traditional risk factors in women; however, mechanisms underlying this
association remain unknown.
Earlier AAM has also been associated with other traditional
cardiometabolic risk factors, including higher BMI,[47] ^5 , [48]^6
higher blood pressure, and abnormal glycemia in adolescents.[49] ^7 It
has been reported that the relationship between AAM and body mass index
(BMI) may mediate the association between AAM and cardiovascular health
outcomes, albeit findings from prospective studies have reported
contradictory results when adjusting for BMI in relation to menarche
onset and cardiometabolic outcomes in adults.[50] ^6 , [51]^7
Similarly, it remains unclear whether systolic blood pressure (SBP) and
glycemic and lipid traits serve as a potential confounder in the AAM
and CVD relationship.[52] ^8 , [53]^9 , [54]^10 Previous studies
consistently demonstrated the potential causal association of AAM with
CVD outcomes, coronary artery disease,[55] ^11 , [56]^12 and ischemic
heart disease.[57] ^13 Several studies have additionally proposed that
AAM is associated with various cardiometabolic risk factors, including
SBP, BMI, blood lipid levels, and markers of subclinical
atherosclerosis (ie, cIMT). All these cardiometabolic traits and
diseases are considered risk factors for future CVD events. However,
most of them have not examined the potential mediatory effect of these
risk factors. Whether the previously reported associations between AAM
and CVD reflect a causal effect and whether they are mediated by BMI
and other cardiometabolic risk factors remains to be determined.
cIMT, a marker of subclinical vascular disease, predicts future
cardiovascular events including myocardial infarction and stroke and
its timely recognition can slow or prevent the progression to overt
CVD. The association between AAM and cIMT has been investigated in only
a few observational studies, with contradictory findings.[58] ^14
Similarly, there is some evidence to suggest an association between
AAM, lipid traits, glycemic traits and blood pressure, although the
evidence is also conflicting.[59] ^15 , [60]^16 A better understanding
of the relationship between pubertal timing, modifiable cardiometabolic
risk factors, and subclinical atherosclerosis may contribute to early
intervention and improved disease management.
The use of genomic analyses such as linkage disequilibrium score (LDSC)
regression and MR may enable us to unravel the complex associations
between cardiometabolic risk factors, AAM, and cIMT. LDSC is a method
that provides unbiased estimates for the genetic correlation between 2
traits.[61] ^17 , [62]^18 MR is an instrumental variable method that
exploits the random assignment of exposure‐associated risk alleles,
enabling us to detect the potentially causal relationships between risk
factors and disease outcomes.[63] ^19 MR framework can overcome
limitations of observational studies by using genetic variants as
instrumental variables that are randomly distributed at conception,
thereby mimicking randomized controlled trials,[64] ^20 and are less
prone to reverse causation, measurement error, and reverse
causality.[65] ^21 Numerous studies have investigated the potential
causal associations of AAM with CVDs, but a vast number of those have
mainly focused on the clinical CVD outcomes including coronary artery
disease, myocardial infarction, and stroke.[66] ^12 , [67]^13 , [68]^22
, [69]^23 No MR study has investigated the causal associations between
AAM and preclinical vascular disease as measured by cIMT.
In this study, we aimed to (1) estimate cross‐trait genetic
correlations and identify genome‐wide genetic overlap between AAM,
cIMT, and modifiable cardiometabolic risk factors; (2) perform
univariable MR analyses to test if genetic liability to AAM causally
increases the risk of subclinical atherosclerosis using cIMT as a
marker; and (3) perform 2‐step MR to investigate whether clinically
relevant cardiometabolic risk factors mediate effects of genetic
liability to AAM on cIMT.
Methods
An overview of the study design is shown in Figure [70]1.
Figure 1. Schematic design showing (A) the main principles of Mendelian
randomization and (B) workflow of the proposed causal relationships between
AAM, cardiometabolic risk factors, and cIMT.
Figure 1
[71]Open in a new tab
Based on the MR assumptions, the genetic variants are assumed to affect
cIMT through AAM only, not through confounding factors. AAM indicates
age at menarche; BMI, body mass index; cIMT, carotid artery
intima‐media thickness; FG, fasting glucose; HDL, high‐density
lipoprotein cholesterol; IVW‐RE, random‐effects inverse
variance‐weighted; LDL, low‐density lipoprotein cholesterol; MR,
Mendelian randomization; MR‐PRESSO, Mendelian Randomization Pleiotropy
Residual Sum and Outlier; PM, proportion mediated; SBP, systolic blood
pressure; SNPs, single nucleotide polymorphism; TC, total cholesterol;
TG, triglycerides; UVMR, univariate Mendelian randomization; and WM,
weighted median.
Genetic Association Data Sources
Age at Menarche
Genetic associations of AAM were obtained from a large‐scale
sex‐specific genome‐wide association study (GWAS) meta‐analysis among
329 345 women of European ancestry (ReproGen consortium [N=179 117],
23andMe [N=76 831], and UK Biobank [N=73 397];[72] ^24 Table [73]S1).
Carotid Intima‐Media Thickness
We performed a sex‐stratified GWAS of cIMT (mean of the maximum, unit
of measurement [μm]) to identify genetic instruments associated with
cIMT in the UK Biobank (UKBB) sample (Figure [74]2, Table [75]S2). A
description of UKBB resources and cIMT phenotyping have been described
in detail previously[76] ^25 , [77]^26 (UKBB Resource 511,
[78]https://biobank.ctsu.ox.ac.uk/crystal/refer.cgi?id=511). In brief,
a GWAS of cIMT was performed on 22 000 participants of European‐only
ancestry (mean age, 56.5 [range, 40–69] years). For the GWAS of cIMT in
the UKBB set, we use natural log‐transformed values of the mean of the
maximum cIMT for normality. We restricted the samples to European‐only
individuals on the basis of the UKBB data field 21 000
([79]https://biobank.ctsu.ox.ac.uk/crystal/field.cgi?id=21000). The
GWAS of cIMT was performed using a linear mixed noninfinitesimal model
as implemented in BOLT‐LMM software.[80] ^27
Figure 2. Genome‐wide association study (A) cIMT and (B) SBP.
Figure 2
[81]Open in a new tab
SNPs with P<0.05 and 1e‐5 were plotted for cIMT and SBP, respectively.
For SBP, SNPs with P<5e‐20 were annotated to nearest gene to avoid
overlap (genome build 37). Red dotted line indicates genome‐wide
threshold of P<5e‐8. cIMT indicates carotid artery intima‐media
thickness; SBP, systolic blood pressure; and SNPs, single‐nucleotide
polymorphisms.
Modifiable Cardiometabolic Risk Factors
Anthropometric Traits
Genetic associations of adult BMI (kg/m^2) were obtained from a UKBB
GWAS among 193 570 female participants
([82]http://www.nealelab.is/uk‐biobank/).[83] ^28 Inverse rank‐normal
transformation was applied to the raw BMI value constructed using
height and weight measured during the initial center visit
(Table [84]S1).
Systolic Blood Pressure
We performed a sex‐specific GWAS of systolic blood pressure (SBP;
mm Hg) using data on up to 230 000 European‐only women participants
(mean age, 56.60 [SD, 7.87] years; mean SBP, 138.0064 [SD, 21.178]
mm Hg) using the UKBB resources (Figure [85]2, Table [86]S3). We
adjusted for medication use by adding 15 mm Hg to SBP measurements for
individuals reported to be taking blood pressure–lowering medication
(N=94 289, UKBB data field 6153
[[87]https://biobank.ndph.ox.ac.uk/ukb/field.cgi?id=6153]). The
sex‐specific GWAS was performed using BOLT‐LMM software,[88] ^27 and
the analysis was adjusted for age and the first 4 principal components
on the basis of the UKBB data.
Glycemic Traits
Instrumental variables associated with fasting glucose (FG) were
extracted from a large‐scale sex‐specific GWAS from the meta‐analyses
of glucose and insulin‐related traits consortium (MAGIC) (n=73 089
women).[89] ^29 Briefly, FG was measured in whole blood and corrected
to plasma level using the correction factor of 1.13, and individuals
were excluded from the analysis if they had a physician diagnosis of
diabetes, were on diabetes treatment (oral or insulin), or had a
fasting plasma glucose ≥7 mmol/L[90] ^29 (Table [91]S1).
Lipids
Instrument variables associated with high‐density lipoprotein
cholesterol (HDL‐C; mmol/L), low‐density lipoprotein cholesterol
(LDL‐C; mmol/L), triglycerides (mmol/L), and total cholesterol (TC;
mmol/L) were retrieved from publicly available GWAS summary data from
the Neale laboratory [UKBB—Neale laboratory;
[92]http://www.nealelab.is/uk‐biobank/].[93] ^28 The analyses included
European‐only participants and were adjusted for age, age2, sex, and
the first 20 principal components.
The studies used in our analysis were approved by their respective
institutional review boards, and informed consent was provided by all
participants. Further details on study cohorts used for the analysis
are provided in Table [94]S1.
Cross‐Trait LDSC
We applied LDSC to quantify the genetic overlap between AAM, BMI, SBP,
HDL‐C, LDL‐C, triglycerides, TC, FG, and cIMT using the LDSC software
with GWAS summary statistics data on up to 1.2 million HapMap
single‐nucleotide polymorphisms (SNPs).[95] ^17 , [96]^18 Briefly, LDSC
estimates the genetic covariance by regressing the cross‐products of
summary test statistics (Z1Z2 statistics) for 2 traits against its
LDSC.[97] ^17 , [98]^18 The slope (coefficient) represents genetic
correlation. We used the precomputed LDSCs of European samples
estimated using the 1000 Genomes reference panel provided on the
website of this software ([99]http://github.com/bulik/ldsc). The
analysis was restricted to HapMap3 SNPs, as these are well‐imputed in
most studies.[100] ^30 The “munge_sumstats.py” and “ldsc.py” from the
LDSC command line tool were used to convert the GWAS summary statistics
for LDSC regression and to conduct LDSC, respectively. To account for
multiple testing, we applied a 5% false discovery rate.
Two‐Sample Univariable MR
We performed a 2‐sample univariable MR to investigate the effect of AAM
on cIMT. Figure [101]1 shows the overall study design that we have
adopted for this analysis. We selected SNPs associated with AAM at P
value <5×10^−8. SNPs with a minor allele frequency≤5% and F‐statistics
<10 were excluded. Correlated SNPs (r2≥0.001) were excluded by keeping
the one with the strongest association with AAM (ie, with the smallest
P value). Genetic associations of cIMT for these SNPs were then
obtained from the GWAS of cIMT.
To further investigate the mediating effects of our selected modifiable
risk factors on the AAM–cIMT association, we applied a 2‐step MR
strategy.[102] ^31 Specifically, we performed 2‐sample univariable MR
analyses for investigating the potential causal effect of AAM on BMI,
SBP, HDL‐C, LDL‐C, triglycerides, TC, FG, and the potential causal
effect of on BMI, SBP, HDL‐C, LDL‐C, triglycerides, TC, and FG on cIMT
to assess the mediating effect. To estimate the proportion mediated, we
used the formula
[MATH:
PM=β1×β2βTE :MATH]
, where
[MATH: β1
:MATH]
represents the causal effect of the exposure on the mediator (step 1),
[MATH: β2
:MATH]
represents the causal effect of the mediator on the outcome (step 2),
and
[MATH: βTE
:MATH]
represents the total causal effect of the exposure on the outcome.[103]
^31 We estimated the 95% CIs for proportion mediated using the delta
method.[104] ^31 We used the same criteria as described previously to
select genetic instruments in each of these MR analyses. Genetic
associations of the outcome of interest in each MR analysis for the
selected genetic instrument were then obtained from the corresponding
GWAS.
SNP‐specific causal effects for the selected genetic instruments were
estimated using the Wald ratio, that is, SNP–outcome association
divided by SNP–exposure association.[105] ^32 To obtain the MR effect
estimates, we pooled the SNP‐specific estimates using inverse‐variance
weighted (IVW) random effects model as the main method,[106] ^33 and we
used weighted median and MR‐Egger regression as sensitivity methods to
assess the robustness of IVW estimates and horizontal pleiotropic
effects.[107] ^34 , [108]^35 We also performed MR‐Lasso, which is an
extension of IVW method by excluding the invalid SNPs that had a
nonzero direct effect on the outcome.[109] ^36 Potential outlier SNPs
identified using Mendelian Randomization Pleiotropy Residual Sum and
Outlier (MR‐PRESSO) were excluded from the analysis.[110] ^37 As AAM
and cIMT had overlapping samples with the selected modifiable risk
factors, we applied a recent statistical method, MRlap, which
calculates corrected causal estimates, accounting for sample overlap,
winner's curse, and weak instrument bias in MR.[111] ^38 To orient the
direction of causality, we applied Steiger filtering, which excludes
the SNPs with a larger variance explained in the outcome than in the
exposure.[112] ^39 We accounted for multiple comparisons of 5 outcome
categories using Bonferroni correction with a P value threshold of 0.01
(0.05/5).
To investigate the heterogeneity, we applied MR‐Clust to identify
genetic variants that may have reflected distinct causal
mechanisms.[113] ^40 In brief, MR‐Clust uses an
expectation‐maximization algorithm to assign genetic variants
contributing to the causal effect estimate in a similar direction,
magnitude, and precision to the same cluster.[114] ^40 In this study,
we assigned a genetic variant to a cluster if the conditional
probability was ≥0.8. Here, we focused only on the substantive
clusters, not the null cluster of genetic variants, which do not
suggest a causal effect or the junk clusters of genetic variants that
cannot be assigned to any substantive clusters.
Next, for each cluster, we performed gene‐based pathway enrichment
analysis using ingenuity pathway analysis
([115]www.qiagen.com/ingenuity) to provide biological insights into the
observed cluster‐specific estimates. SNPs in each cluster that had the
conditional probability ≥0.8 were mapped to genes using the University
of California Santa Cruz Genome Browser queried using mysql.[116] ^41
We used a 20‐kb positional mapping window to assign SNPs to gene for
pathway analysis.
Statistical Analysis
All analyses were performed in R 3.6.1 (R Foundation for Statistical
Computing, Vienna, Austria). GWAS of cIMT was performed using
BOLT‐LMM[117] ^26 using an in‐house pipeline. Genetic correlation
across the traits were computed using the software package LDSC
([118]https://github.com/bulik/ldsc). MR analyses were performed using
the TwoSampleMR ([119]https://mrcieu.github.io/TwoSampleMR/), MR‐PRESSO
([120]https://github.com/rondolab/MR‐PRESSO), MendelianRandomization
([121]https://cran.r‐project.org/web/packages/MendelianRandomization),
mrclust packages ([122]https://github.com/cnfoley/mrclust), and MRlap
([123]https://github.com/n‐mounier/MRlap).
Results
Linkage Disequilibrium Score Regression
Cross‐trait LDSC regression revealed strong inverse genetic correlation
between AAM and BMI (r [g ]=−0.32, SE=0.02, P=2.2×10^−58), SBP (r [g
]=−0.09, SE=0.01, P<0.001), and triglycerides (r [g ]=−0.10, SE=0.01,
P=5.9×10^−8) and positive genetic correlation between AAM and HDL (r [g
]=0.14, SE=0.02, P=2.2×10^−12) after a 5% false discovery rate
(Figure [124]3, Table [125]S4). In addition, we examined phenotypic
correlation and cross‐trait LDSC regression between BMI and SBP. The
analysis indicated a phenotypic correlation of 0.26 (P<0.001) and a
genetic correlation (rg) of 0.21 (P<0.001), suggesting that BMI is
correlated with SBP both phenotypically and genetically.
Figure 3. Genetic correlation estimates between AAM, cardiometabolic risk
factors, and cIMT (A) heatmap of genetic correlation (rg) results and (B)
Forest plot.
Figure 3
[126]Open in a new tab
A, Significant genetic correlations at an false discovery rate of 5%
are marked with an asterisk while the dot shows nominally significant
genetic correlation with raw P value ≤0.05 and false discovery rate
>0.05. BMI indicates body mass index; cIMT, carotid artery intima‐media
thickness; FG, fasting glucose; HDL, high‐density lipoprotein
cholesterol; LDL, low‐density lipoprotein cholesterol; NS, not
significant; SBP, systolic blood pressure; TC, total cholesterol; and
TG, triglycerides.
Association of AAM With cIMT
Our analysis showed that genetically predicted earlier AAM was
associated with higher cIMT using genetic variants passing Steiger
filtering (β[IVW]=−0.008, P [IVW]=0.002 [95% CI, −0.013 to −0.003];
Figure [127]4, [128]Table, Tables [129]S5 through [130]S7). Sensitivity
methods showed consistent direction of effect, and MR‐Egger did not
suggest potential pleiotropy. Neither MR‐Lasso nor MR‐PRESSO identified
invalid or outlier SNPs. The MRlap estimates were consistent with the
IVW estimates (β[MRlap]=−0.109, P [MRlap]=0.009; Table [131]S6,
Figure [132]S1). While not applying Steiger filtering, high
heterogeneity was observed (P [Heterogeneity]=0.022) and SNPs with
potential direct effect on the outcome were identified by MR‐Lasso.
However, MR‐Lasso estimate (excluding invalid SNPs) was consistent with
IVW estimates (β[MR‐Lasso]=−0.010, P [MR‐Lasso]=0.001 [95% CI, −0.015
to −0.004]).
Figure 4. Forest plot of effect estimates from Mendelian randomization
analysis. Line segments represent 95% CIs.
Figure 4
[133]Open in a new tab
Solid symbols indicate associations with a P‐value smaller than or
equal to 0.01. MR‐Lasso is not performed for the effect of AAM/BMI/SBP
on cIMT because no invalid SNPs were identified by MR‐Lasso. BMI
indicates body mass index; cIMT, carotid artery intima‐media thickness;
FG, fasting glucose; HDL, high‐density lipoprotein cholesterol; LDL,
low‐density lipoprotein cholesterol; SBP, systolic blood pressure; TC,
total cholesterol; and TG, triglycerides.
Table 1.
Relationships Between AAM, cIMT, BMI, and SBP From MR With Steiger
Filtering
Exposure Outcome Number of SNPs Method[134]^* Beta SE P value 95% CI
(lower) 95% CI (upper) Heterogeneity P value Pleiotropy P value
Main analysis (AAM–cIMT)
AAM cIMT 183 IVW‐RE −0.008 0.003 0.002 −0.013 −0.003 0.984 NA
AAM cIMT 183 WM −0.008 0.005 0.094 −0.017 0.001 NA NA
AAM cIMT 183 MR‐Egger −0.002 0.009 0.779 −0.020 0.015 0.983 0.502
AAM cIMT 183 MR‐Lasso No invalid SNPs were identified
2‐step MR: (1) AAM–BMI; (2) BMI–cIMT
AAM BMI 176 IVW‐RE −0.109 0.012 3.1E‐20 −0.132 −0.085 3.9E‐30 NA
AAM BMI 176 WM −0.086 0.013 4.2E‐11 −0.112 −0.060 NA NA
AAM BMI 176 MR‐Egger −0.039 0.034 0.253 −0.106 0.028 1.4E‐28 0.031
AAM BMI 115 MR‐Lasso −0.100 0.009 1.9E‐29 −0.117 −0.082 NA NA
BMI cIMT 132 IVW‐RE 0.016 0.005 0.002 0.006 0.025 0.858 NA
BMI cIMT 132 WM 0.011 0.009 0.191 −0.006 0.028 NA NA
BMI cIMT 132 MR‐Egger 0.041 0.018 0.022 0.006 0.076 0.878 0.133
BMI cIMT 132 MR‐Lasso No invalid SNPs were identified
2‐step MR: (1) AAM–SBP; (2) SBP–cIMT
AAM SBP 187 IVW‐RE −0.471 0.166 0.004 −0.795 −0.147 1.8E‐20 NA
AAM SBP 187 WM −0.331 0.193 0.086 −0.709 0.047 NA NA
AAM SBP 187 MR‐Egger 0.342 0.543 0.529 −0.722 1.406 5.8E‐20 0.118
AAM SBP 133 MR‐Lasso −0.362 0.127 0.004 −0.610 −0.114 NA NA
SBP cIMT 217 IVW‐RE 0.002 2.4E‐04 1.3E‐22 0.002 0.003 0.637 NA
SBP cIMT 217 WM 0.003 3.7E‐04 1.1E‐16 0.002 0.004 NA NA
SBP cIMT 217 MR‐Egger 0.005 0.001 4.6E‐11 0.004 0.006 0.863 1.0E‐04
SBP cIMT 217 MR‐Lasso No invalid SNPs were identified
[135]Open in a new tab
AAM indicates age at menarche; BMI, body mass index; cIMT, carotid
intima‐media thickness; IVW‐RE, inverse variance‐weighted random
effects model; MR, Mendelian randomization; MR‐PRESSO, Mendelian
Randomization Pleiotropy Residual Sum and Outlier; NA, not applicable;
SBP, systolic blood pressure; SNPs, single‐nucleotide polymorphisms;
and WM, weighted median.
^*
MR‐Lasso identified invalid SNPs for AAM–BMI and AAM–SBP. For the rest
of the analysis, MR‐Lasso is equivalent to IVW method, thus the results
are not presented to avoid duplication. Potential outliers (MR‐PRESSO)
were only identified for AAM–BMI and AAM–SBP analyses. For these
analyses, we presented the results excluding potential outliers
identified from MR‐PRESSO.
Mediating Effects of BMI and SBP
Genetically predicted earlier AAM was associated with higher adult BMI
(β[IVW]=−0.109, P [IVW]=3.1×10^−20 [95% CI, −0.132 to −0.085];
Figure [136]4, [137]Table, Tables [138]S8 and [139]S9). MR‐Egger
suggested potential pleiotropy (P [Pleiotropy]=0.031); however,
MR‐Lasso estimate (β[MR‐Lasso]=−0.100, P[MR‐Lasso]=1.9×10^−29 [95% CI,
−0.117 to −0.082]; [140]Table) was consistent with the IVW estimate.
MRlap estimates (β[MRlap]=−0.21, P [MRlap]=3×10^−18; Table [141]S9;
Figure [142]S1) showed significant and consistent direction of effect
with the IVW estimate. To further address the heterogeneity and
pleiotropy that we observed between AAM and BMI, we excluded SNPs that
were also associated with BMI at P<0.05. After this exclusion, AAM and
BMI MR estimates remained significant (Table [143]S10; β[IVW]=−0.029, P
[IVW]=6.4×10^−4 [95% CI, −0.04 to −0.012]), and we did not observe any
heterogeneity (P [heterogeneity]=0.55; Table [144]S10) or pleiotropy (P
[Pleiotropy]=0.79; Table [145]S10). Additionally, all MR methods showed
consistent direction of effect (Table [146]S10). Genetically predicted
adult BMI was positively associated with cIMT (β[IVW]=0.016, P
[IVW]=0.002 [95% CI, 0.006–0.025]; [147]Table, Table [148]S11,
Figure [149]S1). All sensitivity methods as well as MRlap that
accounted for sample overlap showed consistent direction of effect, and
MR‐Egger did not suggest potential pleiotropy (Figure [150]S1).
Proportion mediated by adult BMI for the AAM–cIMT association was
estimated to be 20.1% (95% CI, 1.4%–38.9%).
Genetically predicted earlier AAM was also associated with higher SBP
(β[IVW]=−0.471, P [IVW]=0.004 [95% CI, −0.795 to −0.147];
Figure [151]4, [152]Table, Table [153]S5). MR‐Egger estimate
(β[MR‐Egger]=0.342, P[MR‐Egger]=0.529; [154]Table) showed the opposite
direction to the IVW estimate, but MR‐Lasso estimate
(β[MR‐Lasso]=−0.362, P [MR‐Lasso]=0.004 [95% CI, −0.610 to −0.114];
Figure [155]4, [156]Table) was consistent with the IVW estimate.
Genetically predicted adult SBP was positively associated with cIMT
(β[IVW]=0.002, P [IVW]=1.3×10^−22 [95% CI, 0.002–0.003]; Figure [157]4,
[158]Table, Table [159]S11). All sensitivity methods including MRlap
showed consistent direction of effect, but MR‐Egger suggested potential
pleiotropy (P [Pleiotropy]=1.0×10^−4). We next excluded all AAM SNPs
that were also associated with SBP at P<0.05. After applying this
exclusion, IVW and MR‐Lasso were significant (Table [160]S10; P<0.05)
and there was no indication of potential heterogeneity or pleiotropy
(Table [161]S10; P [heterogeneity]>0.1 and P [Pleiotropy]>0.1). The
proportion mediated by adult SBP for the AAM–cIMT association was
estimated to be 13.5% (95% CI, 0.5%–26.6%).
Among the biomarkers under investigation, genetically predicted AAM was
positively associated with TC and HDL (cholesterol: β[IVW]=0.021, P
[IVW]=0.001 [95% CI, 0.005–0.037]; HDL: β[IVW]=0.038, P [IVW]=2.2×10^−4
[95% CI, 0.018–0.059]; [162]Table and Tables [163]S5 and [164]S8) and
with higher triglycerides (β[IVW]=−0.028, P [IVW]=0.002 [95% CI, −0.047
to −0.010]; Figure [165]4, [166]Table and Tables [167]S5 and [168]S8).
Although potential pleiotropy was not suggested, MR‐Egger showed
consistent direction with the IVW method for the above associations
except for triglycerides. However, among these biomarkers, we found
that genetically predicted cholesterol was positively associated with
cIMT (β[IVW]=0.013, P [IVW]=7.6 × 10^−4 [95% CI, 0.005–0.021];
Figure [169]4, [170]Table and Table [171]S11). In addition, genetically
predicted LDL was also positively associated with cIMT (β[IVW]=0.017, P
[IVW]=1.1 × 10^−5 [95% CI, 0.010–0.026]; Figure [172]4, [173]Table and
Table [174]S11). In cases where MR‐Lasso identified outlier SNPs for
each AAM and cardiometabolic risk factor MR, we performed a subsequent
IVW analysis after excluding these outliers. The resulting estimates
consistently displayed significant effects in a consistent direction
with the main analysis (Table [175]S8). The results from MRlap were
consistent with the primary IVW method for all modifiable risk factors
tested (Table [176]S9).
Cluster‐Specific MR Estimates
We found substantial heterogeneity in the AAM–BMI and AAM–SBP analyses
even after excluding invalid SNPs identified using Steiger filtering
and MR‐PRESSO ([177]Table and Table [178]S5). Next, we performed
cluster‐specific MR using MR‐Clust to test whether the clusters had
different causal effects on BMI and SBP than the overall causal effect
using all SNPs. Cluster‐specific SNPs were used as IVs to run the
analysis. Using MR‐Clust,[179] ^40 we identified 3 clusters for AAM–BMI
analysis and 3 clusters for AAM–SBP analysis (Figures [180]5 through
[181]7 and Table [182]S12). All 3 cluster‐specific MR estimates for
AAM–BMI varied in terms of number of SNPs and showed inverse causal
association with BMI. Cluster_1 (N[SNP]=7, β=−0.15, P=2 × 10^−64;
Figures [183]5 through [184]7 and Table [185]S12) showed less strong
causal association as compared with clusters 2 and 3 (Cluster_2
[N[SNP]=22, β=−0.41, P=2×10^−239]), Cluster_3 (N[SNP]=29, β=−0.42,
P<0.001; Figures [186]5 through [187]7 and Table [188]S12). We found
substantial heterogeneity in cluster‐specific estimates for SBP
(Figures [189]5 and [190]7 and Table [191]S12). Cluster_1 (NSNP=5,
β= 3.93, P=3.6× 10^−08), Cluster_3 (NSNP=6, β=2.27, P=8.4×10^−08)
showed positive causal estimates between AAM‐SBP while Cluster_2
(NSNP=24, β=−4.34, P=3.9×10^−79) showed an inverse causal AAM–SBP
association (Figures [192]5 and [193]7 and Table [194]S12).
Figure 5. Forest plot of effect estimates from cluster specific Mendelian
randomization performed using MR‐Clust.
Figure 5
[195]Open in a new tab
Line segments represent 95% CIs. BMI indicates body mass index; cIMT,
carotid artery intima‐media thickness; FG, fasting glucose; HDL,
high‐density lipoprotein cholesterol; LDL, low‐density lipoprotein
cholesterol; SBP, systolic blood pressure; TC (mmol/L), total
cholesterol; and TG (mmol/L), triglycerides.
Figure 7. Clustering of genetic variants for the Mendelian randomization
analysis of age at menarche with adult systolic blood pressure.
Figure 7
[196]Open in a new tab
To provide further insights into the cluster‐specific results, we next
performed ingenuity pathway analysis ([197]www.qiagen.com/ingenuity) to
identify canonical pathways associated to SNPs in each of the clusters.
We performed ingenuity pathway anal[198] yses by grouping SNPs that had
an inclusion probability >80% for each cluster originating from
MR‐Clust (Figures [199]5 through [200]7 and Table [201]S13). Genes from
3 clusters associated with BMI were mapped to distinct canonical
pathways with BAG2 Signaling Pathway (Cluster_1), Uracil Degradation II
(Reductive) (Cluster_2), and Methylglyoxal Degradation III (Cluster_3)
showing the highest enrichment (Table [202]S14). Genes from 3 clusters
associated with SBP were mapped to Chronic Myeloid Leukemia Signaling,
Senescence Pathway, and Endocannabinoid Cancer Inhibition Pathway
(Cluster_3), Thrombopoietin Signaling (Cluster_2) (Table [203]S15).
Figure 6. Clustering of genetic variants for the Mendelian randomization
analysis of age at menarche with adult body mass index.
Figure 6
[204]Open in a new tab
Discussion
Here, we investigated the SNP‐based genetic correlation and potential
causal relationship between AAM and cIMT using LDSC and the MR
paradigm, respectively. To the best of our knowledge, this is the first
study to report genetic correlation and causal associations between AAM
and cIMT, and to investigate whether the effect of AAM on cIMT is
mediated by modifiable risk factors including anthropometric, glycemic,
lipid traits, and blood pressure. First, our cross‐trait LDSC
identified an inverse genetic correlation between AAM, BMI, and cIMT,
suggesting that shared genetic variants and interconnected biological
pathways predispose to these traits. Second, our univariable MR
revealed that genetic liability to earlier AAM is potentially causally
deleterious to cIMT. Finally, using a 2‐step MR framework, we found
that BMI and SBP account for as much as 20.1% and 13.5% of the causal
effect of AAM on cIMT, respectively. Previous MR studies consistently
demonstrated the potential causal association of AAM with CVD outcomes,
including coronary artery disease[205] ^11 , [206]^12 and ischemic
heart disease.[207] ^13 Several studies have additionally proposed that
AAM is associated with various cardiometabolic risk factors, including
SBP, BMI, blood lipid levels, and markers of subclinical
atherosclerosis (ie, cIMT). All these cardiometabolic traits/diseases
are considered risk factors for future CVD events. Therefore, it is
reasonable to claim that implementing early preventive interventions to
address these potential risk factors could be effective in preventing
CVD events. Our study contributes valuable findings in this regard by
demonstrating that one third of the effect of AAM on cIMT is mediated
by modifiable CVD risk factors, namely, BMI and SBP.
Genetic Correlations Among the Traits Under Investigation
First, we investigated the genetic overlap between AAM, cIMT, and
modifiable cardiometabolic risk factors and found that AAM showed
stronger and significant inverse genetic correlation with BMI, and
triglycerides than cIMT and other cardiometabolic risk factors
investigated. Our findings are consistent with previous studies,[208]
^24 , [209]^42 and suggest that AAM and BMI may share genetic variants
that may implicate shared biological pathways between both traits. The
genetic correlation between AAM, cIMT, and SBP were smaller as compared
with BMI and nonsignificant, suggesting that these traits play a
smaller role in the underlying biology of AAM than BMI. To test whether
the observed correlations could reflect causal associations, we used MR
to identify potential causal links between these traits. Our findings
from cross‐trait LDSC were mostly concordant with our MR findings.
Despite the genetic association and MR effect that we identified
between AAM and cIMT, the absence of evidence for genetic correlation
between the traits is noteworthy. MR employs a selection of genetic
variants, as opposed to LDSC analysis, which may reflect an association
that was otherwise canceled out by opposing effects in a genome‐wide
correlation. The genome‐wide genetic correlation between 2 traits can
be influenced by multiple common causes and may be negligible in cases
where some genetic loci exhibit a positive genetic correlation between
the 2 traits, while others show a negative genetic correlation.
Increased intima‐media thickness is a marker of subclinical
atherosclerosis and medial hypertrophy, which has been shown to predict
future cardiovascular events including myocardial infarction and
stroke. Several conventional observational studies have investigated
the inverse association of AAM with cardiovascular risk factors,
including BMI, blood pressure, and cIMT.[210] ^43 , [211]^44 Although
observational and MR studies have consistently shown an inverse
association between AAM and BMI and SBP, the association between AAM
and cIMT is still controversial. For example, a cross‐sectional study
failed to identify any association between AAM <12 years and cIMT,[212]
^45 while a community‐based study reported an inverse and independent
association between early AAM and internal cIMT in Black women.[213]
^46 Additionally, those studies were primarily focused on examining the
association between AMM and cIMT and did not examine the mediating
effects of BMI and SBP. Our 2‐step MR analysis indicated that BMI and
SBP could serve as potential mediators on the effect of AAM on cIMT.
Effect of AAM on cIMT via BMI
Our finding of the mediating role of BMI on the effect of an earlier
AAM on higher cIMT was consistent with existing knowledge. The
relationship between BMI and cIMT is established, with a number of
studies reporting that higher childhood BMI is associated with thicker
cIMT.[214] ^47 , [215]^48 Previous studies have also shown that
children with obesity are more likely to have an early AAM, with fat
distribution affecting this relationship.[216] ^44 , [217]^49 Few
studies have explored the relationship between increased adiposity and
cIMT in adults. A previous MR analysis indicated that earlier AAM
causes higher adult BMI.[218] ^50 Although the exact biological
mechanisms underlying the observed mediation are yet to be fully
discovered, there could be several plausible biological mechanisms for
explaining the relationships. As previously hypothesized,[219] ^50 it
is possible that early AAM leads to an altered hormonal state and
subsequent changes in psychological state and physiological or
metabolic rates of girls at adolescence, leading to weight gain and
associated complications of higher adiposity later in life. This
remains to be investigated in future studies.
Effect of AAM on SBP and cIMT
There have been inconsistent findings regarding the relationship
between AAM and SBP in prior epidemiological studies, with a variety of
studies showing that AAM is inversely associated with
hypertension,[220] ^51 , [221]^52 while others show that AAM is
positively associated with hypertension.[222] ^51 , [223]^53 , [224]^54
Several cross‐sectional studies have suggested that SBP is linearly
associated with cIMT in individuals with and without hypertension.[225]
^55 On the other hand, and in line with our findings, Fan et al
identified an inverse causal relationship between AAM and SBP, with SBP
mediating the effect of AAM on CAD using the MR framework on an Asian
population nested in the Taiwan Biobank.[226] ^12 Other observational
studies in Asian population groups have identified a linear
relationship between increased SBP and cIMT.[227] ^56 Overall, these
findings provide supporting evidence for the mediatory role of SBP in
the effect of earlier AAM on cIMT and promote interventions aimed at
lowering SBP to reduce risk of cIMT thickening attributable to earlier
AAM and its clinical sequelae. As suggested by a recent article, early
AAM is a feature of early vascular aging in adults with primary
hypertension, and primary hypertension is known to cause vascular
damage that may increase cIMT and CVD risk.[228] ^57
Effect of AAM on Modifiable Cardiometabolic Risk Factors
Our MR analyses identified an inverse association between AAM and
triglycerides and positive association between AAM and HDL levels.
These associations have been previously reported in a MR study
investigating the association of genetically predicted AAM and
myocardial infarction.[229] ^23 Associations between dyslipidemia and
earlier AAM have also been highlighted in traditional epidemiological
studies.[230] ^43 Additionally, previous studies have highlighted
significant effects of genetically predicted AAM on fasting blood
glucose levels using GWAS summary data from sex‐combined data sets for
blood glucose.[231] ^58 , [232]^59 Our study did not replicate this
association using sex‐specific GWAS estimates for exposure and outcome
data sets. This discrepancy in findings highlights the need to increase
the generation and use of sex‐specific GWAS in downstream MR analyses
to identify possible sex‐specific causal pathways in relevant disease
phenotypes. Overall, our findings suggest genetically predicted earlier
AAM have a potentially causal effect on increased cIMT through BMI and
adiposity‐related mechanisms that are also reflected through adverse
lipid profiles.
Strengths and Limitations
Our study has a number of strengths:
1. We leverage state‐of‐the‐art MR methodology to examine the potential
causal cardiometabolic effects of earlier AAM, an experiment that is
not practically and ethically suitable in a randomized controlled trial
setting.
2. The GWAS data and genetic instruments used in the cross‐trait LDSC
regression and MR analyses were derived from the largest
female‐specific GWAS of AAM, cIMT, BMI, SBP, lipids traits, and FG.
3. We used multiple sensitivity methods to improve robustness of model
assumptions and our analysis showed consistent direction of effect
estimates supporting the validity of our findings.
Our study has several limitations:
1. The sex‐specific GWAS of AAM, BMI, cIMT, SBP, and lipid traits were
derived from UKBB resources that partly overlapped. While overlapping
samples are not known to produce any systematic bias for the genetic
correlation analysis, it could potentially attenuate the MR findings.
Previous studies have shown that sample overlap in MR can increase type
1 error and could potentially lead to bias results for classic MR
methods.[233] ^60 We performed sensitivity analysis using the recently
proposed MRlap method to address the sample overlap.[234] ^38 Our
findings remain consistent; however, we observed that, for some traits,
the effect size was larger, with wider CIs. Further replication of our
results in independent data sets is warranted.
2. Our 2‐step MR was based on the assumption that there was no exposure
× mediator interaction.[235] ^31 The presence of exposure × mediator
interaction is known to produce biased estimates from both 2‐step and
multivariable MR. However, the size of the bias is thought to be lower
in the MR analyses as compared with the non‐instrumental variable based
methods.[236] ^31
3. The GWAS resources drawn in the current study were drawn from the
European‐only women participants of UKBB. Therefore, our findings may
not be applicable to women of other populations and ethnicities, as
allelic differences between ethnicities may produce different effect
estimates.
4. We did not include childhood BMI in the mediation analysis as the
GWAS summary level data available from the Genetic Investigation of
Anthropometric Traits consortium had a relatively small sample size.
5. Our choice to employ 2‐step MR instead of multivariable MR to
investigate the mediating effects of BMI and SBP was influenced by
several factors. First, we were primarily interested in estimating the
mediating effect of BMI and SBP individually and the potential impact
of intervening on these mediators. Second, the power of MR analyses is
reported to decrease as the number of mediators increases in the
multivariable MR model.[237] ^31 Although MR mediation analysis has
advantages compared with non‐instrumental variable regression‐based
methods and counterfactual approaches, which are still susceptible to
issues such as unmeasured confounding and reverse causality in
conventional observational studies, MR mediation analysis has
limitations as well. For example, MR mediation analysis assumes a
linear association and does not account for exposure–mediator
interaction.[238] ^31 Finally, MR relies on core assumptions that
impact the reliability of causal inference. We aimed to ensure the
robustness of our findings through the implementation of multiple
sensitivity analyses and filters on genetic instrument selection to
overcome potential pleiotropic effects. Causal inferences made through
our MR analyses should be interpreted carefully, as MR estimates are
only suggestive of potential effects of genetic predisposition of an
exposure on an outcome of interest but do not provide evidence that
therapeutic intervention targeting the exposure will change the
outcome.
Clinical and Public Health Implications
Our findings suggest that early interventions targeting SBP and BMI may
result in significant health benefits by reducing cIMT and its
associated adverse health outcomes in women. Despite the growing
availability of sex‐specific data focused on cardiovascular health
care, the implementation and translation of findings stemming from such
data in practical settings has been slow. Given the lack of progress in
improving the incidence and mortality rates of CVD among younger women,
our results motivate further research aimed at enhancing our
understanding of female cardiovascular health and risk factors. This
knowledge will be crucial to advance the global prevention, diagnosis,
and treatment of CVD in women.
Conclusions
Our findings suggest that an earlier AAM is potentially causally
associated with higher cIMT and that one third of the effect of AAM on
cIMT may be mediated by the modifiable risk factors BMI and SBP. Those
findings may indicate that early intervention aimed at lowering BMI and
hypertension may be beneficial in reducing the risk of developing
subclinical atherosclerosis due to earlier age at menarche. A better
understanding of the relationship between reproductive factors and CVD
may contribute to better prevention and management of this disease.
Sources of Funding
The authors acknowledge support from the Imperial College British Heart
Foundation Centre for Research Excellence (RE/18/4/34215), the UK
Dementia Research Institute at Imperial College London (MC_PC_17114),
and National Institute for Health and Care Research Imperial Biomedical
Research Centre.
Disclosures
None.
Supporting information
Tables S1–S15
Figure S1
[239]JAH3-13-e032192-s001.pdf^ (1.2MB, pdf)
Acknowledgments