Association of female reproductive traits with altered aging trajectories: Insights from genetic and observational analyses

Summary

Women’s reproductive health plays a pivotal role in both longevity and the aging process. We conduct Mendelian randomization (MR) and observational analyses to investigate these relationships. Univariate MR analyses reveal that older age at first birth, later menarche, higher estradiol, and sex hormone-binding globulin (SHBG) increase longevity, while pre-eclampsia liability decreases longevity. Older ages at first birth and at first sexual intercourse are associated with lower DNAmGrimAgeAccel, but these associations disappear after mutual adjustment. Mediation analyses identify cardiometabolic diseases, lung diseases, and mental disorders as key mediators. In corroborating the MR results, observational analyses show that early reproductive behaviors, such as age at first sex, are associated with accelerated biological aging. Additionally, we observe significant non-linear associations between hormone levels, age at menopause, and aging outcomes. This study highlights the impact of reproductive health on aging and suggests potential strategies for promoting healthy aging in women.

Graphical abstract

Highlights

• •Female reproductive traits causally shape longevity and biological aging
• •Early reproductive behaviors are associated with accelerated biological aging
• •Cardiometabolic, lung, and mental disorders mediate the reproductive-aging links
• •Hormone levels and menopause show non-linear patterns with biological aging

Reproductive health strongly shapes women’s longevity and biological aging. Using genetic and observational approaches, Lou et al. link reproductive behaviors, hormones, and diseases to aging outcomes, highlighting potential strategies to promote healthy aging in women.

Introduction

Reproductive health is a cornerstone of women’s overall well-being and has profound implications for aging and longevity.^1,2 Despite women generally living longer than men, they experience higher rates of frailty and health complications in later life,³ partly due to the significant physiological demands of reproduction, including menstrual bleeding, pregnancy, fluctuations of sex hormones, and reproductive disorders.^4,5,6,7 Reproductive factors have been associated with telomere length, indicating the intricate connection between reproductive health and aging phenotypes.⁸ These findings underscore the need for further investigation into how female reproductive traits influence aging outcomes, with implications for promoting healthier aging in women.

Aging is a complex biological process that gradually impairs cellular, tissue, and organ function,^9,10 involving epigenetic changes that increase disease susceptibility and mortality risk.¹¹ Individuals with longer lifespans often exhibit a biological age younger than their chronological age, emphasizing the link between biological aging and longevity.¹² Biological age can be estimated through DNA methylation (DNAm) at specific cytosine-guanine dinucleotide (CpG) sites or by algorithms combining clinical biomarkers.^13,14,15 Biological age acceleration (AgeAccel) is defined as the residual difference between biological and chronological age, reflecting the degree of discrepancy between the two. Evidence suggests that female reproductive traits may influence aging phenotypes; however, studies have yielded mixed results. Gravidity has been reported to promote epigenetic age acceleration in both young women and postmenopausal individuals.^16,17 Additionally, studies have identified associations between epigenetic clocks, fertility,¹⁸ and age at menarche (AAM).¹⁹ Although one previous Mendelian randomization (MR) study has explored sexual development and aging, its reliance on mixed-sex data for reproductive behaviors has introduced inaccuracies.²⁰ Moreover, the mechanisms mediating the relationship between female reproductive health and aging remain poorly understood. Recent studies indicate that reproductive health in women is closely related to a variety of diseases affecting multiple organs,^21,22,23,24 many of which may be preventable. Therefore, it is plausible that certain conditions mediate the impact of female reproductive health on aging.

This study aimed to systematically investigate the relationship between female reproductive traits and aging. We employed two-sample univariable MR (UVMR) and multivariable MR (MVMR) methods to assess the total and independent effects of reproductive traits on aging phenotypes while minimizing confounding and reverse causality.²⁵ Subsequently, we explored the mediating roles of disease categories and hormones using a two-step MR framework.²⁶ Additionally, an observational study was carried out using data from the National Health and Nutrition Examination Survey (NHANES), allowing for the exploration of both linear and non-linear relationships between female reproduction and aging, providing deeper insights into this complex interplay.

Results

Characteristics of the MR study

Figure 1 outlines the study design overview of the three-part MR and observational studies. Genome-wide association study (GWAS) data sources for 37 female reproductive traits (reproductive behaviors, diseases, and hormones) and 6 aging phenotypes (percentile longevity and four DNAm-based measures of biological age acceleration), as well as 40 potential mediators, are summarized in Tables S1–S3.

Figure 1.

Overview of the study design

This study investigates the genetic causality and observational associations between female reproductive traits and aging phenotypes. First, UVMR was applied to assess the causal links between reproductive traits and aging. Next, we explored the effects of related reproductive traits on aging after mutual adjustment using MVMR. Third, a two-step MR was employed to investigate potential mediating effects. Additionally, an observational study utilizing NHANES data examined both linear and non-linear relationships between reproductive traits and aging. The icons were created using BioRender (https://biorender.com/).

Total effects of reproductive traits on aging phenotypes

Genetically determined per year increase in age at first birth (AFB) (inverse-variance weighted [IVW]-estimated odds ratio [OR], 1.21; 95% confidence interval [CI], 1.12 to 1.30; p = 1.28 × 10⁻⁶) and per log-transformed unit increase in estradiol levels (OR, 1.35; 95% CI, 1.29 to 1.41; p = 4.02 × 10⁻³⁶) were causally associated with higher odds of achieving 90th percentile longevity, both results remaining significant at the phenotype-adjusted Bonferroni threshold (p value < 2.25 × 10⁻⁴, Figure 2A). Conversely, genetic liability to pre-eclampsia was strongly associated with lower odds of reaching 90th percentile longevity. Genetically predicted older age at first sexual intercourse (AFS), AAM, and elevated sex hormone-binding globulin (SHBG) levels were suggestively associated with increased odds of achieving 90th percentile longevity, while genetically predicted levels of follicle-stimulating hormone (FSH) had the opposite effect (nominally significant, p < 0.05). However, reproductive behaviors were not significantly associated with 99th percentile longevity. For this extreme aging measure, genetically predicted levels of estradiol (OR, 1.44; 1.24 to 1.68; p = 1.84 × 10⁻⁶) and SHBG (OR, 1.41; 1.01 to 1.98; p = 0.045) maintained positive effects (Figure 2B).

Figure 2.

UVMR estimates of the causal relationships between female reproductive traits and aging phenotypes

(A–F) The female reproductive traits are categorized into three groups: orange for hormones, blue for diseases, and red for behaviors. UVMR estimates used the IVW method, with positive, negative, and non-significant effects represented in red, blue, and gray, respectively. Dots indicate beta coefficients or ORs with 95% CI error bars. The arrows indicate values exceeding the plot boundaries. Statistical significance was defined at three levels: nominal (∗p < 0.05), trait-adjusted Bonferroni (∗∗p < 0.05/37 reproductive traits = 1.35 × 10⁻³), and phenotype-adjusted Bonferroni (∗∗∗p < 0.05/37/6 aging phenotypes = 2.25 × 10⁻⁴). Asterisks denote the corresponding thresholds. AAM, age at menarche; ALB, age at last live birth; ANM, age at natural menopause; NLB, number of live births; BAT, bioavailable testosterone; FSH, follicle-stimulating hormone; LH, luteinizing hormone; AMH, anti-müllerian hormone; UL, leiomyoma of uterus; EVP, excessive vomiting in pregnancy; GH, gestational hypertension; PCOS, polycystic ovarian syndrome; POF, primary ovarian failure; Longevity90, 90th percentile longevity; Longevity99, 99th percentile longevity; IEAA, intrinsic epigenetic age acceleration; ORs, odd ratios.

Accelerated aging was evaluated through epigenetic AgeAccel, using four distinct DNAm aging clocks (Figures 2C–2F). The most advanced measure, DNAmGrimAgeAccel, which predicts mortality, showed the strongest associations. A genetically determined decrease in AFS (IVW estimated β = −1.05 years; 95% CI: −1.47 to −0.62; p = 1.11 × 10⁻⁶; Figure 2D) and AFB (β = −0.28 years; 95% CI: −0.41 to −0.14; p = 4.39 × 10⁻⁵) was significantly associated with greater DNAmGrimAgeAccel at the phenotype-adjusted Bonferroni threshold. Genetically predicted AFS and SHBG levels also showed negative associations with DNAmPhenoAgeAccel (−0.61; −1.16 to −0.07; p = 0.03) and DNAmGrimAgeAccel (−0.39; −0.72 to −0.05; p = 0.02), respectively. For reproductive disorders, genetic predisposition to excessive vomiting during pregnancy (EVP) and irregular menses were associated with increased epigenetic age acceleration, while pre-eclampsia was associated with decreases (Figure 2). The instrumental validity assessment demonstrated robust strength (lowest F-statistic = 21.85), and significant results after phenotype-adjusted Bonferroni correction were validated by at least one sensitivity analysis (Tables S4 and S5).

Independent effects of female reproductive traits on aging phenotypes

We next conducted mutual adjustments within the three categories of UVMR-positive reproductive traits to account for the complex intercorrelations among these traits. After adjusting for AFS and AFB, genetically predicted AAM remained independently associated with higher odds of achieving 90th percentile longevity (MV-IVW-estimated OR, 1.13; 95% CI, 1.02 to 1.25; p = 0.02) (Figure 3A; Table S6). Similarly, genetically predicted pre-eclampsia was associated with lower odds after adjusting for post-term delivery (0.59; 0.42 to 0.82; p = 1.80 × 10⁻³) (Figure 3A). Genetically predicted SHBG levels remained positively related with higher odds of both the 90th (1.45; 1.10 to 1.92; p = 9.79 × 10⁻³) (Figure 3A) and 99th (1.72; 1.13 to 2.62; p = 1.11 × 10⁻²) (Figure 3B) percentile longevity following the respective adjustments. In terms of epigenetic aging, neither AFB nor AFS showed significant associations with DNAmGrimAgeAccel following mutual adjustments, a result that remained consistent or in the analysis with further inclusion of AAM (Table S6). Additionally, among the five reproductive disorders initially associated with DNAmGrimAgeAccel in the UVMR analysis, none remained associated after mutual adjustments. However, when retaining only the three traits that passed phenotype-adjusted Bonferroni threshold, genetic liability to pre-eclampsia was independently associated with slower DNAmGrimAgeAccel (MV-IVW-estimated β, −0.54; 95% CI, −0.97 to −0.10; p = 1.64 × 10⁻²) (Figure 3C; Table S6). A similar association was observed between genetic liability to pre-eclampsia and DNAmHannumAgeAccel (Figure 3D; Table S6). For DNAmPhenoAgeAccel, mutual adjustments among the reproductive diseases that were significant in UVMR resulted in none remaining associated with this aging measure (Table S6).

Figure 3.

Impact of female reproductive traits on aging after mutual adjustment and mediation effects in their causal associations

(A–D) The MVMR analyses evaluated the independent causal effects of correlated reproductive traits on aging phenotypes, with at least one significant association. MVMR estimates used the IVW method, with positive, negative, and non-significant effects represented in red, blue, and gray, respectively. Dots indicate beta coefficients or ORs with 95% CI error bars. The arrows indicate values exceeding the plot boundaries. Statistical significance is denoted by asterisks, where ∗p < 0.05.

(E) The Sankey diagram illustrates the flow from exposure (female reproductive traits, left) through mediators (center) to aging outcomes (right). The left-side connections display the UVMR estimates for the relationship between the exposure and mediator, while the right-side connections show the MVMR estimates of the causal effects of each mediator on aging, adjusted for the corresponding reproductive traits. Red lines show positive associations, and blue lines negative, with dark blue marking heart failure as the main mediator. Only exposure-mediator and mediator-outcome effects with consistent directions are shown. p < 0.05 was considered significant.

(F) Forest plot showing exposure-mediator effects, mediator-outcome effects in MVMR, indirect effects, and mediation proportions, estimated using the Delta method. Only exposure-mediator and mediator-outcome effects with consistent directions are shown. CHD, coronary heart disease; T2D, type 2 diabetes.

Additionally, we assessed the relationship between aging phenotypes and a combination of the most significant traits obtained from each reproductive trait category. After adjusting for estradiol and pre-eclampsia, genetically predicted AFB was independently associated with higher odds of 90th percentile longevity (OR, 1.20; 95% CI, 1.08 to 1.33; p = 5.12 × 10⁻⁴) (Figure 3A). In the mutually adjusted model for DNAmGrimAgeAccel, genetically predicted SHBG and AFS showed a comparable negative correlation with the UVMR result (Figure 3C). All results were validated by at least one sensitivity analysis (Tables S6 and S7).

Potential mediators between female reproductive traits and aging phenotypes

The mediation analysis followed a systematic four-step screening process. Initially, 65 causal relationships were identified between 25 potential mediators and 6 aging phenotypes (Tables S8 and S9). The mediators were required to retain a significant independent causal relationship with aging phenotypes after adjusting for the corresponding reproductive traits that were significant in the UVMR, which reduced the number of mediators to 23 (Tables S14 and S15). Next, reproductive traits had to be causally associated with the mediators, but not vice versa, narrowing the pool to 18 mediators (Table S10. UVMR estimates for the causal associations between female reproductive traits and mediators, related to Figure 3, Table S11. Heterogeneity statistics and horizontal pleiotropy for UVMR analysis between female reproductive traits and mediators, related to Figure 3, Table S12. UVMR estimates for the causal associations between mediators and female reproductive traits, related to Figure 3, Table S13. Heterogeneity statistics and horizontal pleiotropy for UVMR analysis between mediators and female reproductive traits, related to Figure 3). Finally, only fifteen mediators remained, as both the total and mediating effects were in the same direction with mediation proportions greater than zero (Table S16; Figure 3E). Several mediators were identified between female reproductive behaviors and 90th percentile longevity, with a primary focus on cardiometabolic diseases, chronic lung diseases, and mental disorders. Among cardiovascular diseases, heart failure mediated the largest proportion of the association between AFB and 90th percentile longevity (28.16%; 95% CI, 12.98%–46.81%), followed by coronary heart disease (CHD), myocardial infarction (MI), stroke, and atrial fibrillation. Alzheimer’s disease (14.81%; 95% CI, 2.98%–27.40%) and chronic obstructive pulmonary disease (COPD) (8.16%; 95% CI, 0.06%–18.88%) were also significant mediators. Ferritin, associated with trace element disorders, accounted for 8.46% of the association. Heart failure also mediated a significant proportion of the association between AFS and longevity (38.06%; 95% CI, 14.50%–67.71%). Other diseases, including atrial fibrillation, type 2 diabetes (T2D), MI, CHD, COPD, and stroke, mediated 8.47%–29.74% of the total effect, with T2D and stroke additionally mediating the association between AFS and DNAm-based biological age acceleration. For AAM, heart failure mediated the largest proportion of its association with 90th percentile longevity, followed by CHD, MI, and COPD. Regarding reproductive hormones, the association between SHBG and 90th/99th percentile longevity was mediated primarily by stroke and Lewy body dementia, accounting for 35.84%–59.86% of the total effects. Lung diseases—such as COPD, lung cancer, lung adenocarcinoma, and lung squamous cell carcinoma—primarily mediated the associations between AFB, AFS, preterm delivery, and DNAm-based biological age acceleration, accounting for 11.88%–44.07% of the total effect. The mediators of the relationships between primary ovarian failure, irregular menses, EVP, post-term delivery, and aging phenotypes were MI, stroke, and T2D (Figure 3F).

Characteristics of the observational study

We collected clinical information of up to 25,059 non-pregnant female participants aged over 20 years old from the NHANES dataset, spanning from 1999 to 2018 (mean [SE] age at baseline, 48.12 [0.20] years). By the end of the follow-up period on December 31, 2019, a total of 3,422 deaths had been recorded, with a median follow-up time of 9.08 years. Table S17 outlines the mean values with standard errors or the distributions of reproductive traits and biological aging markers across the available years. Details of all covariates from 1999 to 2018 are provided in Table S18.

Association between female reproductive traits and longevity

In models 1 and 2, after adjusting for demographic, socioeconomic, and common disease factors, significant associations emerged between age at menopause, various reproductive hormones, and all-cause mortality (Tables S19 and S20). In the full adjusted model, each 1-year increase in both age at menopause (Hazard ratio [HR], 0.988; 95% CI, 0.983 to 0.994; p = 4.26 × 10⁻⁵) and natural menopause (ANM) (0.984; 0.971 to 0.998; p = 0.0255) was associated with reduced fatality risk. Conversely, elevated luteinizing hormone (LH) levels (1.011; 1.003 to 1.019; p = 0.0043) and SHBG levels (1.005; 1.001 to 1.009; p = 0.0151) were associated with higher mortality risk. Older ALB (0.990; 0.982 to 0.998; p = 0.0163) and larger number of live births (NLB) (0.973; 0.951 to 0.995; p = 0.0163) were linked to reduced mortality risk. Infertility was also associated with a lower risk of mortality (Table S21).

A non-linear relationship with higher all-cause mortality was observed as NLB increased from 2.1 to 3.9, while the overall trend exhibited an approximately linear decline in the fully adjusted model (Figure S1). A partially U-shaped like non-linear association was observed between testosterone levels and all-cause mortality. Lower bioavailable testosterone levels (less than 10.1 ng/dL) was linked to higher mortality risk, while, in higher concentrations, the 95% CI encompassed log HR of 0. Total testosterone level associated with the lowest mortality risk was 13.9 ng/dL, with levels above this threshold showing a plateau effect (Figure 4).

Figure 4.

Non-linear associations between female androgen-related markers and all-cause mortality

The pink lines in the chart illustrate the multivariable-adjusted log hazard ratios, with the light pink shaded areas representing 95% CIs derived from restricted cubic spline regressions. Dashed lines at a log hazard ratio of 0 serve as reference lines, positioned at the first change point. All models were adjusted for age, race and ethnicity, poverty-to-income ratio (PIR), educational level, marital status, alcohol intake, body mass index (BMI) category, smoking status, hypertension, diabetes, cardiovascular disease, cancer, history of ovary removal, pregnancy history, menopause status, female hormone use, live birth category, and history of hysterectomy. The model for bioavailable testosterone (BAT) was adjusted for multicollinearity; refer to the method details section for details. CP, change point.

Association between female reproductive traits and DNAm-based biological age acceleration

We assessed biological aging in NHANES using genome-wide DNAm data, incorporating both first- and second-generation epigenetic-aging clocks. In model 1, higher AFS, AFB, and age at menopause were inversely correlated with different clocks. Higher FSH levels were correlated with reduced DNAmPhenoAgeAccel (Table S19). No significant associations were found in model 2 (Table S20). In model 3, which further adjusted for reproductive factors, each additional year of AFS was consistently associated with a decrease in 4 DNAm-based biological age acceleration (β range: −0.345 to −0.261 years; p ≤ 0.0364) (Tables 1 and S21). Older AFB was associated with an increase in intrinsic epigenetic age acceleration (IEAA) (β = 0.119 years; 95% CI, 0.009 to 0.229; p = 0.0349) and DNAmHorvathAgeAccel (β = 0.118 years; 0.007 to 0.229; p = 0.0377). Uterine leiomyoma was also significantly linked to elevated DNAmGrimAgeAccel (2.029 years; 0.086 to 3.972; p = 0.0413).

Table 1.

Associations of the reproductive traits with biological aging in NHANES

Exposure	DNAmHannumAgeAccel		DNAmPhenoAgeAccel		DNAmGrimAgeAccel		IEAA
	p value	Beta (95% CI)	p value	Beta (95% CI)	p value	Beta (95% CI)	p value	Beta (95% CI)
AAM	0.4595	0.096 (−0.165, 0.356)	0.718	−0.067 (−0.441, 0.307)	0.9791	0.002 (−0.137, 0.141)	0.1918	0.170 (−0.090, 0.431)
AFB	0.9157	0.007 (−0.130, 0.145)	0.8472	0.011 (−0.105, 0.127)	0.221	−0.037 (−0.098, 0.024)	0.0349	0.119 (0.009, 0.229)
AFS	0.0364	−0.291 (−0.562, −0.020)	0.0341	−0.345 (−0.662, −0.028)	0.0006	−0.261 (−0.400, −0.122)	0.3607	−0.144 (−0.461, 0.173)
Age at menopause	0.9963	−0.000 (−0.052, 0.052)	0.773	0.011 (−0.067, 0.090)	0.3265	−0.017 (−0.053, 0.018)	0.2892	−0.028 (−0.080, 0.025)
ALB	0.378	−0.037 (−0.123, 0.048)	0.8528	−0.010 (−0.124, 0.103)	0.9674	0.001 (−0.039, 0.040)	0.232	0.039 (−0.026, 0.103)
ANM	0.2965	0.067 (−0.062, 0.196)	0.419	−0.070 (−0.244, 0.104)	0.192	−0.069 (−0.175, 0.037)	0.5517	−0.044 (−0.194, 0.106)
Years ovulating	0.8847	−0.004 (−0.058, 0.050)	0.4124	0.029 (−0.042, 0.100)	0.7723	−0.004 (−0.035, 0.026)	0.3185	−0.026 (−0.078, 0.026)
NLB	0.393	0.099 (−0.135, 0.333)	0.9966	0.000 (−0.236, 0.237)	0.3442	0.055 (−0.062, 0.171)	0.9269	−0.009 (−0.217, 0.198)
PLN	0.0977	−0.419 (−0.919, 0.082)	0.3323	−0.294 (−0.903, 0.316)	0.5896	−0.077 (−0.367, 0.212)	0.7395	−0.050 (−0.357, 0.256)
FSH	0.8912	0.001 (−0.021, 0.023)	0.0539	−0.029 (−0.059, 0.001)	0.5338	−0.005 (−0.023, 0.012)	0.5687	−0.005 (−0.022, 0.012)
LH	0.4795	0.011 (−0.021, 0.043)	0.5406	−0.013 (−0.056, 0.030)	0.6154	0.007 (−0.022, 0.036)	0.6601	−0.007 (−0.038, 0.025)
UL
No	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.
Yes	0.3608	0.866 (−1.045, 2.776)	0.2791	1.376 (−1.180, 3.933)	0.0413	2.029 (0.086, 3.972)	0.7751	−0.310 (−2.518, 1.897)
Preterm delivery
No	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.
Yes	0.2895	1.186 (−1.073, 3.445)	0.1001	2.437 (−0.504, 5.377)	0.808	−0.189 (−1.776, 1.398)	0.6981	0.400 (−1.703, 2.502)
Endometriosis
No	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.	Ref.
Yes	0.4803	1.203 (−2.246, 4.651)	0.6723	−1.151 (−6.674, 4.372)	0.6348	−0.743 (−3.915, 2.429)	0.5645	−0.882 (−3.986, 2.221)

All models were adjusted for age, race and ethnicity, PIR, educational level, marital status, alcohol intake, BMI category, smoking status, hypertension, diabetes, cardiovascular disease, cancer, history of ovary removal, pregnancy history, menopause status, female hormone use, live birth category, and history of hysterectomy. Some models were adjusted for variable duplication; see the method details section for details. p value < 0.05 was considered significant. AAM, age at menarche; ALB, age at last live birth; ANM, age at natural menopause; NLB, number of live births; PLN, number of pregnancy losses; FSH, follicle-stimulating hormone; LH, luteinizing hormone; UL, leiomyoma of uterus.

A complex, non-linear relationship existed between reproductive traits and DNAm-based biological age acceleration. For example, AFB below 24.1 years was associated with lower DNAmGrimAgeAccel (Figure 5A), but the 95% CI for DNAmPhenoAgeAccel crossed 0 (Figure S2). Similarly, for ALB, higher ALB below approximately 31.6 years was correlated with lower DNAmHannumAgeAccel, but wide 95% CIs for other epigenetic clocks suggested weak association (Figure S3). Years ovulating associated with the highest DNAmGrimAgeAccel were 25.9 years, displaying an inverted U-shaped relationship (Figure 5B). However, the remaining aging models showed inconsistent trends (Figure S4). Furthermore, an increase in NLB beyond 3.5 was associated with a significant rise in DNAmGrimAgeAccel (Figures 5C and S5). The relationship between age at menopause and DNAm-based biological age acceleration demonstrated a non-linear pattern beyond age 40, characterized by either negative correlations or fluctuations with wide 95% confidence intervals (Figures 5D and S6). When specifically examining natural menopause across the 40–60 years age range, multiple DNAm-based biological age acceleration measures revealed more nuanced yet similarly non-linear associations. Specifically, earlier ANM below approximately 47–48 years was associated with increased DNAm-based age acceleration. For older ANM, we observed a shallow nadir around 47–48 years and a brief elevation near 50 years, with 95% CIs widening or crossing 0 (Figures 5E–5H and S7).

Figure 5.

Non-linear associations between female reproductive traits and DNA methylation-based biological age acceleration

(A–D) Non-linear associations between female behaviors and DNA methylation-based biological age acceleration.

(E–H) Non-linear associations between ANM and DNA methylation-based biological age acceleration.

(I–L) Non-linear associations between female gonadotropins and DNA methylation-based biological age acceleration. The pink lines represent multivariable-adjusted predicted values, with light pink areas indicating the 95% CIs derived from restricted cubic spline regressions. Dashed lines at a predicted value of 0 serve as reference lines, positioned at the first change point. All models were adjusted for age, race and ethnicity, PIR, educational level, marital status, alcohol intake, BMI category, smoking status, hypertension, diabetes, cardiovascular disease, cancer, history of ovary removal, pregnancy history, menopause status, female hormone use, live birth category, and history of hysterectomy. Some models were adjusted for variable duplication; refer to the method details section for details. CP, change point.

Gonadotropins, including FSH and LH, displayed U-shaped associations with DNAm-based biological age acceleration. FSH concentrations associated with the lowest DNAmGrimAgeAccel, DNAmPhenoAgeAccel, and IEAA were 44.2, 56.5, and 60.9, respectively (Figures 5I–5K). A similar U-shaped association was observed between FSH and other clocks (Figure S8). In terms of LH, concentration associated with the lowest DNAmGrimAgeAccel was 37 IU/L (Figure 5L), while other epigenetic clocks indicated an optimal concentration around 20 IU/L followed by a non-significant plateau phase (Figure S9).

Association between female reproductive traits and biomarker-based biological age acceleration

We then used blood biochemical biomarkers to predict biological aging for other female reproductive traits lacking DNAm-based biological age acceleration data in the relevant years. In the fully adjusted model, no linear associations were found between infertility, gestational diabetes, androgen-related markers, and biomarker-based biological age acceleration (Table S21). However, a distinct non-linear association emerged between androgen-related markers and BioKDMAgeAccel and BioPhenoAgeAccel, where the lowest age acceleration was at a concentration of SHBG of approximately 92.1 nmol/L and 96 nmol/L, respectively (Figure 6). The higher levels of SHBG exhibited a wide 95% CI. Total testosterone levels associated with the lowest BioKDMAgeAccel were 30 ng/dL, while levels below 32 ng/dL were similarly inversely associated with BioPhenoAgeAccel (Figure S10). Additionally, the lowest BioPhenoAgeAccel was at a concentration of bioavailable testosterone of 6.5 ng/dL (Figure 6). We further validated the associations between female reproductive traits and biomarker-based biological age acceleration. In the fully adjusted model, higher AFS was associated with lower BioPhenoAgeAccel. Other reproductive behaviors showed broadly linear associations with BioPhenoAgeAccel, consistent with the non-linear analyses (Table S21; Figure S11). Most reproductive behaviors were not associated with BioKDMAgeAccel, though higher AAM and AFS exhibited a non-linear increase (Figures S12 and S13). The non-linear association between age at menopause and BioKDMAgeAccel fluctuated (Figure S14). The LH concentration associated with the lowest BioPhenoAgeAccel was 38.8 IU/L (Figure S15).

Figure 6.

Non-linear associations between female androgen-related markers and biomarker-based biological age acceleration

The pink lines in the chart represent the multivariable-adjusted predicted value, with the light pink shaded areas indicating the 95% CIs derived from restricted cubic spline regressions. Dashed lines at a predicted value of 0 serve as reference lines, positioned at the first change point. All models were adjusted for age, race and ethnicity, PIR, educational level, marital status, alcohol intake, BMI category, smoking status, hypertension, diabetes, cardiovascular disease, cancer, history of ovary removal, pregnancy history, menopause status, female hormone use, live birth category, and history of hysterectomy. CP, change point.

Discussion

In this comprehensive study, we found that genetically determined later reproductive events, such as older AFB, AFS, and AAM, as well as elevated estradiol and SHBG levels, were causally associated with increased longevity and decelerated biological aging. Conversely, reproductive disorders like pre-eclampsia were associated with reduced longevity and accelerated aging. Cardiometabolic diseases, lung diseases, and mental disorders were identified as major mediators in the association between reproductive traits and aging phenotypes. Observational analyses confirmed the MR results on AFS and characterized the non-linear associations of reproductive hormone levels (testosterone, FSH, and LH) and age at natural menopause with aging outcomes. Our study provides compelling evidence of associations between female reproductive traits and aging phenotypes.

Previous studies have highlighted associations between specific reproductive behaviors and longevity, with higher polygenic scores for earlier reproduction and alleles that promoted reproductive success indicative of lower survival probability²⁷ and menarche at ≥12 years correlating to modestly increased odds of surviving to age 90.²⁸ Our MR analysis indicated that older AAM, AFB, and AFS were causally associated with reaching the 90th percentile of longevity, but these associations did not extend to extreme longevity (99th percentile) or all-cause mortality in the observational study. Altogether, our findings suggest that reproductive behaviors may not directly correlate with lifespan endpoints, potentially under the influence of socioeconomic factors, selection biases,²⁹ and non-linear relationships.³⁰ Substantial evidence connects reproductive behaviors to biological aging, with AFB negatively correlated and birth number weakly associated with epigenetic age acceleration in the Sister Study cohort.³¹ These findings partially align with our MR analysis, where higher AFB and AFS were associated with lower DNAmGrimAgeAccel, supporting their positive impact on longevity, consistent with observational results showing AFS correlated with reduced DNAm-based biological age acceleration across four measures. However, the causal associations were attenuated after mutual adjustment, suggesting that their effects are not independent but may be largely explained by correlations with each other and with other reproductive traits, such as AAM. Mediation analysis revealed that earlier sexual activity, childbirth, and earlier menarche were related with an increased risk of cardiovascular, pulmonary, and psychological diseases later in life, particularly heart failure,³² as supported by various studies.^33,34,35 Premature reproduction may hasten aging through multiple biological pathways. Beyond hormonal perturbations, metabolic dysregulation, and low-grade inflammation that elevate cardiovascular and respiratory vulnerability,^36,37 early pregnancies can induce placental maldevelopment and endothelial dysfunction (vasospasm/activation), mechanistically overlapping with later cardiometabolic and pulmonary tissue remodeling.³⁸ Earlier sexual activity and menarche also increase allostatic load and disturb the crosstalk between the hypothalamic pituitary adrenal (HPA) and gonadal (HPG) axes, fostering insulin resistance, adiposity, and persistent pro-inflammatory signaling that promote arterial stiffness and atherogenesis.^39,40 Extended lifetime sex-steroid exposure and stress-sensitive neurodevelopmental trajectories heighten risks of depression and anxiety, which further worsen cardiometabolic profiles.⁴¹ Epigenetic programming and microbiome-gut-brain interactions may embed these early stressors, sustaining cardiometabolic and psychological vulnerability across the life course.^42,43 These conditions heighten aging and mortality risk, consistent with the antagonistic pleiotropy hypothesis of reproduction and lifespan.²⁷

In observational analyses, ANM showed a suggestive, model-dependent non-linear association with DNAm-based biological age acceleration: a shallow nadir around 47–48 years and a short-lived elevation near ∼50 years. We interpret this as time-sensitive physiology at the final menstrual period (FMP)—abrupt estradiol withdrawal with a rise in FSH and transitional shifts in inflammation, cardiometabolic profiles, vascular/autonomic tone, and sleep-related stressors—rather than a durable effect of later ANM per se.^44,45,46 Second-generation DNAm clocks (e.g., GrimAge) integrate smoking- and protein-based surrogates linked to these systems and may be especially responsive during the peri-/early postmenopausal window. Earlier ANM occurring before ∼47 was associated with higher DNAm age acceleration in our non-linear models, consistent with reports by Levine et al.,⁴⁷ potentially reflecting lower lifelong endogenous estrogen and downstream vascular and immune dysregulation.^48,49 By contrast, results above ∼47 were largely null, which helps reconcile the absence of an overall linear causal effect in MR analyses based on genetically proxied ANM. Linear MR is not designed to detect segment-specific or transient dynamics, and epigenetic-aging GWASs used for MR include both sexes, which may dilute female-specific transition signals.⁵⁰ Notably, gene-expression studies indicated marked shifts around 49–51 coincident with pronounced physiological change across the transition.⁵¹ Taken together, ANM is unlikely to exert a large, persistent linear effect on epigenetic aging, but proximity to the FMP—especially younger ANM—may transiently influence DNAm-based biological aging readouts and deserves longitudinal, non-linear causal follow-up.

Pregnancy-related complications were identified as being associated with aging phenotypes and changes in biological age before and after pregnancy. Poganik et al. found that, from the onset to the later stages of pregnancy, the biological age of the pregnant woman increases.⁵² Our study identified pregnancy-related vomiting and gestational hypertension as contributors to accelerated biological aging, which may partially explain prior findings that pregnancy promotes aging. However, Pham et al. demonstrated that transitioning from pregnancy to the non-pregnant state triggers a “recovery” or even a “reversal,” wherein the reduction in maternal biological age exceeds the gain from gestational increase.⁵³ Intriguingly, our study also identified an association between genetically determined pre-eclampsia and diminished DNAm-based biological age acceleration, suggesting a potential manifestation of this “reversal.” Nevertheless, consistent with findings from a national Swedish cohort study,⁵⁴ we observed that pre-eclampsia correlates with lower longevity. While we propose that pregnancy-induced immune regulation may be temporarily protective against aging by decelerating changes in aging biomarkers, it is also possible that DNAm clocks do not fully capture the vascular damage-driven aging processes that underlie pre-eclampsia,⁵⁵ and thus the apparent biological age deceleration could be an artifact of measurement limitations.⁵⁶ This suggests differing short- and long-term associations of these conditions.

Evidence on reproductive hormones and aging is limited, with most focusing on men,^57,58 whereas our study fills this gap by examining these associations in women. While higher SHBG levels were found to increase longevity and slow down biological aging in MR, we disclosed a non-linear relationship between biological aging, SHBG, and androgen, with low levels accelerating aging. Notably, MR and observational analyses yielded divergent findings for SHBG, which may be explained by differences in endpoints (longevity vs. all-cause mortality), population background, or residual pleiotropy and confounding. These discrepancies highlight the need for cautious interpretation of SHBG’s role in aging. A negative association between DNAmPhenoAgeAccel and SHBG in 1,062 postmenopausal women partially supports our finding of the non-linear relationship.⁵⁹ This could be explained by our mediation analysis, which demonstrated that elevated SHBG levels were associated with reduced risks of both stroke and Lewy body dementia, supporting its beneficial role in promoting longevity. These findings may reflect the protective effects of SHBG and androgen on cardiometabolic diseases at specific level,^60,61 particularly major adverse cardiovascular events (MACEs). For instance, total testosterone levels in the highest quartile (22.8 ng/dL) compared to the lowest quartile (4.9 ng/dL) were associated with a reduced incidence of MACEs in women over 70,⁶² which aligns with our observation that low testosterone levels (below ∼24 ng/dL) were correlated with increased biological aging and mortality risk. These findings suggest a potential non-linear relationship in which low testosterone levels may be harmful, whereas higher levels may offer no additional benefit and could even be associated with risks, although evidence for a precise threshold is still limited. The underlying mechanism may involve that excessive testosterone can induce vascular constriction and platelet aggregation while also contributing to visceral fat deposition and increasing the risk of hypertension and insulin resistance.^63,64 Our findings revealed that hormonal benchmarks were associated with minimal aging and mortality, providing valuable indicators for assessing female aging in clinical practice.

Apparent discrepancies between MR and observational findings reflect differences in design, populations, power, and measurement. MR treats exposures as quasi-static instruments and is less confounded, whereas smaller, multiethnic cohorts are prone to residual confounding, assay/batch variation, limited methylation waves, and survival/publication biases. DNAm clocks (used in MR GWAS) capture molecular aging, while biomarker clocks index contemporaneous physiology. Moreover, biological aging markers can diverge from mortality outcomes. For example, higher parity is linked to lower mortality, supported by the Dubbo and the Irish Longitudinal Study on Ageing (TILDA) studies, likely reflecting selection bias: healthier women both bear more children and live longer.^65,66 In contrast, our non-linear models show that high parity (NLB > 3.5) associates with higher DNAmGrimAgeAccel, consistent with the transient inflammatory/metabolic burden of repeated pregnancies, whereas moderate parity (<3.5) shows weak elevation, explaining the lack of a linear trend. Critically, mortality integrates lifelong exposures and multiple aging risks, while DNAm clocks provide a state-dependent snapshot of molecular dysregulation at sampling; divergence between these outcomes is therefore expected. Regarding age at menopause, DNAmGrimAge2’s added CRP/HbA1c surrogates heighten sensitivity to perimenopausal inflammatory-metabolic shifts, producing non-linear variability, whereas DNAmGrimAge shows greater stability with non-linear associations. Biomarker clocks index proximal, labile physiology. BioKDMAgeAccel reflects short-term “functional age,” whereas BioPhenoAgeAccel integrates biomarkers calibrated to mortality risk; these construct differences plausibly underlie divergence from MR and DNAm clocks. Accordingly, BioPhenoAgeAccel largely aligns with MR, while KDMAgeAccel shows a non-linear AFS effect (slower aging before ∼19.8 years, then plateau), and, for age at menopause, both KDMAgeAccel and DNAmGrimAgeAccel exhibit non-linearity consistent with perimenopausal volatility around the FMP.

A key strength of this study is its comprehensive focus on women, integrating MR and observational analyses to explore links between various female reproductive phenotypes and aging. We included three aspects of female reproductive traits, capturing a comprehensive view of reproduction, along with diverse aging phenotypes and numerous potential mediators. This approach enables a thorough and multifaceted analysis. The use of NHANES data allowed us to generalize these findings to a broader and more diverse population beyond those of European descent and enabled us to explore potential non-linear relationships. Moreover, by utilizing a robust MR framework with strict selection criteria for mediators, we ensured the validity of causal inferences. Our primary findings were further supported by multiple sensitivity analyses, consistently demonstrating the consistency and dependability of the findings.

In conclusion, our study revealed the intricate relationship between female reproductive traits and aging, suggesting the importance of prioritizing female reproductive health in the context of population aging. Our findings suggest that strengthening screening and management of cardiometabolic diseases and mental disorders in women with a history of early reproductive events, reproductive disorders, or hormonal dysregulation may represent effective public health strategies to mitigate accelerated aging and promote healthy aging.

Limitations of the study

This study also has several limitations. First, GWASs of longevity, biological aging, and FSH/LH included both sexes, although our observational analyses focused on women. Some exposures (e.g., estradiol, FSH, and EVP) were instrumented by few SNPs; despite F > 10, results warrant cautious interpretation. Second, menopausal status was adjusted for, but cycle-related hormonal fluctuations could not be fully addressed, and some variables such as EVP were unavailable in observational study. Third, DNAm data were limited to two waves, focusing on women aged 50 and above, primarily postmenopausal, with a relatively small participant number. Fourth, the two-step MR mediation analysis assumes a unidirectional pathway, and horizontal pleiotropy may bias estimates. Finally, ANM interpretation is constrained by cross-sectional DNAm data, recall bias, and unmodeled time since menopause, while MR yields only average linear effects, limiting detection of female-specific or non-linear dynamics.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Liangshan Mu (mu.liangshan@zs-hospital.sh.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

This study is based on existing, publicly available datasets. The GWAS summary statistics used in this study are openly accessible, with detailed sources provided in Tables S1–S3. Data from the National Health and Nutrition Examination Survey (NHANES) can be obtained online at https://wwwn.cdc.gov/nchs/nhanes/. This paper does not report original code. Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.

Acknowledgments

This study was supported by the National Natural Science Foundation of China (82288102) and the Beijing Natural Science Foundation (Z230013). We sincerely appreciate the researchers who made the GWAS data available to the public, and we extend our thanks to all the participants involved in the GWAS studies mentioned in this manuscript. Icons in the graphical abstract were created using BioRender (https://biorender.com/).

Author contributions

Conceptualization, C.L., G.W., L.M., S.Y., and Y.Z.; data curation, C.L., G.W., and L.M.; formal analysis, C.L. and G.W.; funding acquisition, T.Y.L., S.Y., Y.Z., and L.M.; investigation, C.L. and G.W.; methodology, C.L. and G.W.; project administration, T.Y.L., S.Y., Y.Z., and L.M.; resources, C.L., G.W., Z.X., Y.C.J., and Y.L.; software, C.L., G.W., M.Z., H.Y., J.W., and W.Z.; supervision, H.Y., L.W., L.H., H.-M.C., and X.D.; validation, G.W.; visualization, C.L.; writing – original draft, C.L. and G.W.; writing – review and editing, T.Y.L., S.Y., Y.Z., and L.M.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Age at first birth GWAS	Mills et al.67	https://www.ebi.ac.uk/gwas/studies/GCST90000048
Age at first sex GWAS	Mills et al.67	https://www.ebi.ac.uk/gwas/studies/GCST90000045
Age at last live birth GWAS	D’Urso et al.68	https://cloudstor.aarnet.edu.au/plus/s/RkRnbpX5WPU3QI3
Years ovulating GWAS	D’Urso et al.68	https://cloudstor.aarnet.edu.au/plus/s/RkRnbpX5WPU3QI3
Number of live birth GWAS	UK biobank	https://broad-ukb-sumstats-us-east-1.s3.amazonaws.com/round2/additive-tsvs/2734.gwas.imputed_v3.female.tsv.bgz
Gestational duration GWAS	Solé-Navais et al.69	https://egg-consortium.org/Gestational-duration-2023.html
Age at menarche GWAS	Day et al.70	https://www.reprogen.org/Menarche_1KG_NatGen2017_WebsiteUpload.zip
Age at natural menopause GWAS	Ruth et al.71	https://www.reprogen.org/reprogen_ANM_201K_170621.txt.gz
Preterm delivery GWAS	Solé-Navais et al.69	https://egg-consortium.org/Gestational-duration-2023.html
Post term delivery GWAS	Solé-Navais et al.69	https://egg-consortium.org/Gestational-duration-2023.html
Reproductive diseases GWAS	FinnGen R10	https://storage.googleapis.com/finngen-public-data-r10/summary_stats
Anti-Müllerian hormone GWAS	Verdiesen et al.72	https://www.ebi.ac.uk/gwas/labs/studies/GCST90104596
Estradiol GWAS	Schmitz et al.73	https://www.ebi.ac.uk/gwas/labs/studies/GCST90020092
Bioavailable testosterone GWAS	Ruth et al.74	https://www.ebi.ac.uk/gwas/labs/studies/GCST90012102
Sex hormone-binding globulin GWAS	Ruth et al.74	https://www.ebi.ac.uk/gwas/labs/studies/GCST90012106
Total testosterone GWAS	Ruth et al.74	https://www.ebi.ac.uk/gwas/labs/studies/GCST90012112
Follicle-stimulating hormone GWAS	INTERVAL	https://gwas.mrcieu.ac.uk/datasets/prot-a-528/
Luteinizing hormone GWAS	INTERVAL	https://gwas.mrcieu.ac.uk/datasets/prot-a-529/
Prolactin GWAS	INTERVAL	https://gwas.mrcieu.ac.uk/datasets/prot-a-2375/
Progesterone GWAS	Pott et al.75	https://zenodo.org/records/5644896/files/Progesterone.zip?download=1
DNAm biological age acceleration GWAS	McCartney et al.76	https://www.ebi.ac.uk/gwas/studies
Percentile longevity GWAS	Deelen et al.77	https://www.ebi.ac.uk/gwas/studies
Observational data	NHANES	https://wwwn.cdc.gov/nchs/nhanes/
Software and algorithms
R TwoSampleMR package	Github	https://mrcieu.github.io/TwoSampleMR/
R MendelianRandomization package	CRAN	https://cran.r-project.org/web/packages/MendelianRandomization/index.html
R version 4.2.2	R Foundation for Statistical Computing, Vienna, Austria	https://www.r-project.org/
R MR-PRESSO package	Github Verbanck et al.78	https://github.com/rondolab/MR-PRESSO
R survival package	CRAN	https://cran.r-project.org/package=survival
R survey package	CRAN	https://cran.r-project.org/package=survey
R rms package	CRAN	https://cran.r-project.org/package=rms
R MVMR package	Github	https://github.com/WSpiller/MVMR
R RMediation package	CRAN	https://cran.r-project.org/web/packages/RMediation/index.html

Experimental model and study participant details

GWAS sources for exposures

We included three categories of female reproductive traits, all derived from European female populations: reproductive behaviors, reproductive diseases, and reproductive hormones. The reproductive behaviors category includes age at first birth (AFB), age at first sex (AFS),⁶⁷ age at last live birth (ALB), years ovulating,⁶⁸ number of live births (NLB), gestational duration,⁶⁹ age at menarche,⁷⁰ and age at natural menopause.⁷¹

The reproductive diseases category is further divided into pregnancy-related complications, such as preterm delivery, post-term delivery,⁶⁹ gestational hypertension, gestational diabetes, pre-eclampsia, excessive vomiting in pregnancy, spontaneous abortion, habitual abortion, and eclampsia; uterine and ovarian disorders, including endometriosis, primary ovarian failure, ovarian dysfunction, uterine leiomyoma, adenomyosis, menorrhagia, irregular menses, and oligomenorrhoea; and reproductive system conditions, such as infertility, ectopic pregnancy, and polycystic ovarian syndrome. The reproductive hormones category includes anti-Müllerian hormone (AMH),⁷² estradiol (Estradiol was examined as a binary trait, divided into two groups: one with estradiol levels exceeding the detection limit and the other with levels falling below it.),⁷³ follicle-stimulating hormone (FSH), prolactin, luteinizing hormone (LH),⁷⁹ progesterone,⁷⁵ and androgen-related markers,⁷⁴ including total testosterone, bioavailable testosterone (BAT), and sex hormone-binding globulin (SHBG). Details of each exposure are provided in Table S1. All public GWAS datasets have been carefully quality-controlled, as outlined in their original publications.

GWAS sources for mediators

Based on a comprehensive literature review,⁸⁰ we identified 40 traits potentially acting as mediators in the pathways connecting female reproductive health to aging. These traits span categories include cancer, mental disorders, muscle health, hormonal balance, cardiometabolic diseases, chronic lung diseases, vitamin D and iron metabolism disorders, and female hormones. To pinpoint mediators of the causal link between female reproductive traits and aging phenotypes, we employed four rigorous criteria. First, the mediator must have a causal connection with aging phenotypes. Second, the mediator need to have a causal effect on aging phenotypes, independent of the associated female reproductive traits. Third, female reproductive traits should causally influence the mediator, with no evidence of reverse causation. Fourth, the total effect and the mediating effect should be in the same direction. Detailed information about each mediator is provided in Table S3.

GWAS sources for outcomes

The study examined aging phenotypes, focusing on both longevity and biological aging. For longevity, we assessed the genetics of individuals who exceeded the average life expectancy for their demographic group, based on population-specific and gender-specific mortality rates. GWAS used life tables stratified by country, gender, and birth year to identify ages corresponding to different survival thresholds. This approach distinguished between “cases”—individuals who had achieved exceptional longevity—and “controls,” who had not. Two meta-analyses of GWAS, utilizing a stringent definition of longevity, were included in our study. These analyses comprised 11,262 cases who survived at or beyond the 90th survival percentile and 3,484 cases who survived at or beyond the 99th survival percentile. 25,483 controls were included, with their age at death or last follow-up was at or below the age matching the 60th percentile of survival.⁷⁷

For biological aging, the study utilized epigenetic clocks to estimate biological aging through DNAm data. We employed four epigenetic clocks from both first- and second-generation models. The first-generation clocks estimate chronological age, whereas the second-generation clocks assess multisystem phenotypic age and predict mortality by integrating additional health-related factors. Together, these models provide a more comprehensive assessment of aging by capturing a wider spectrum of biological aging processes.⁸¹

The first generation of epigenetic clocks included the Hannum Clock (DNAmHannumAge),⁸² which predicted chronological age using 71 CpG sites selected through elastic net regression, and the Horvath Clock,⁸³ which estimated chronological age based on 353 CpG sites across 51 tissues. The Horvath Clock provided a measure of intrinsic epigenetic age acceleration (IEAA) that accounts for differences in immune cell types. The second generation of epigenetic clocks includes DNAmPhenoAge,⁸⁴ which predicts “Phenotypic Age”—an indicator of aging-related mortality risk—using 513 CpG sites. DNAmGrimAge,⁸⁵ the most advanced clock, predicts all-cause mortality by integrating age, sex, and 1,030 CpG sites associated with health factors like smoking and leptin levels. Both GWAS and meta-analyses were performed using the calculated biological ages, comparing them to chronological age to derive age acceleration metrics.⁷⁶ Details of each outcome are provided in Table S2.

Observational study population

The NHANES is an ongoing national cross-sectional survey, with data publicly available through the U.S. Centers for Disease Control and Prevention (CDC) Website (http://www.cdc.gov/nchs/nhanes.htm). The study protocol received approval from the Research Ethics Review Board of the National Center for Health Statistics (NCHS), and all participants provided written informed consent during recruitment. In this study, we analyzed data from 1999 to 2018, focusing on women aged 20 years and older, excluding pregnant participants, with a maximum sample size of 25,059. The study adhered to the STROBE guidelines,⁸⁶ as detailed in Table S23.

Ethics statement

The summary-level GWAS data used in this analysis are publicly available, thereby exempting this study from ethical review requirements. Ethical approvals for the original GWASs are documented in the respective studies referenced in this manuscript. Furthermore, the NHANES study received ethical approval from the National Center for Health Statistics Research Ethics Review Board, ensuring compliance with ethical standards for participant data usage.

Method details

MR analysis study design

All GWAS and observational data employed in this study were publicly accessible. Ethical approvals for the referenced GWASs were documented in the original publications cited within this manuscript. The MR study, in three phases (Figure 1), began with Phase 1 assessing the causal links between female reproductive traits and aging phenotypes using two sample univariable MR (UVMR). Phase 2 evaluated whether these causal effects were independent among reproductive traits with strong associations by applying multivariable MR (MVMR). Phase 3 investigated potential mediators in the causal relationship between female reproductive traits and aging phenotypes, quantifying the mediating effects of each candidate through two-step MR approach. The study follows the STROBE-MR guidelines (Table S22).⁸⁷

UVMR and MVMR analysis

All MR analyses adhered to three fundamental assumptions. First, in UVMR, genetic variants must show a strong association with the exposure; in MVMR, they should be significantly related to at least one exposure. Second, genetic variants should not be connected to confounders that could affect the relationship between the exposure instruments and the outcomes. Third, genetic variants should influence outcomes only through their effects on the exposure, with no direct causal pathway.²⁵ We initially utilized UVMR to evaluate the causal effect of each female reproductive traits and mediator on aging outcomes. Then, we used MVMR to explore the independent causal effects of each reproductive trait on aging phenotypes. The female reproductive traits included in MVMR were selected based on a positive UVMR outcome. Two selection criteria were applied. First, mutual adjustments were made within the three categories of reproductive traits, provided that the category contained more than one trait with a significant UVMR association. Second, mutual adjustments were made for the three most significant traits in each category.

For UVMR analysis, genetic instruments were chosen based on genome-wide significance (p < 5 × 10⁻⁸) while applying a linkage disequilibrium filter (r² < 0.001, kb = 10,000). If fewer than two independent SNPs were available, the significance threshold was relaxed to p < 5 × 10⁻⁶.⁸⁸ The LD reference panel was sourced from the European super-population in the 1000 Genomes Project, concentrating on bi-allelic SNPs with a minor allele frequency exceeding 1%. For unavailable SNPs in the outcome summary statistics, proxy SNPs in high LD (r² > 0.8) were identified using the LDproxy.⁸⁹ Exposure and outcome were harmonized to guarantee that the SNP estimates reflected the same effect allele. Allele frequency data was employed to verify the alignment of alleles between the exposure and outcome GWAS. Palindromic SNPs with a minor allele frequency greater than 0.42 were excluded from the analysis.⁹⁰ For MVMR, SNPs associated with all exposures were selected as instrumental variables. The combined SNP set was clumped using LD (r² < 0.001) within a 10,000 kb window, based on the lowest p-values for either trait. The harmonization procedure for the exposure and outcome datasets mirrored the one used in UVMR.

We applied the random-effects inverse-variance weighted (IVW) method as the main approach for UVMR analysis.⁹¹ For multivariable MR (MVMR) analysis, we used the multivariable IVW (MV-IVW) method.⁹² For exposures with one instrumental variable, the Wald ratio method was employed in UVMR. The IVW method effectively combines variant-specific ratio estimates and accounts for heterogeneity in causal estimates from individual variants,⁹³ while the MV-IVW method incorporates correlations among multiple exposures.⁹²

To confirm the reliability of our results, we conducted multiple sensitivity analyses. IV strength was assessed by F-statistics; F > 10 indicates low weak-instrument bias. For each SNP, F = R²(N − k − 1)/[k(1 − R²)], where R² = β²/(β² + SE² × N), with β as the effect size, SE as the standard error, N as the sample size, and k as the number of IVs. For each exposure–outcome pair, the overall F-statistic equals the sum of SNP-specific F-statistics.⁹⁴ Heterogeneity was evaluated with Cochran’s Q statistics.⁹⁴ For UVMR, we corroborated the robustness of the IVW estimates with additional methods, including the weighted median,⁹⁵ weighted mode,⁹⁶ MR-Egger,⁹⁷ and MR pleiotropy residual sum and outlier (MR-PRESSO) methods.⁷⁸ For MVMR, we confirmed the the stability of the MV-IVW estimates using multivariable MR-Egger (MVMR-Egger)⁹⁸ and multivariable MR-Lasso (MVMR-Lasso) methods.⁹²

MR mediation analysis

We investigated the intermediary role of particular traits in the connection between female reproductive characteristics and aging through a two-step MR analysis.²⁶ In the initial step, we employed UVMR to evaluate the causal impact (β₁) of female reproductive traits on potential mediators. To address potential bidirectionality that could confound the mediation model, we conducted reverse MR analysis to evaluate the causal relationships between mediators and reproductive traits. Any mediators exhibiting reverse causality were excluded from further analysis. Subsequently, for mediators that demonstrated causal associations with aging in UVMR, we proceeded with MVMR to estimate the causal effect (β₂) of these mediators on aging, while adjusting for the corresponding female reproductive traits. The mediation proportion for each factor was calculated by multiplying β₁ by β₂ and dividing the results by the total effect of female reproductive traits on aging phenotypes. Standard errors and 95% confidence intervals for these mediation proportions were estimated using both the delta method^26,99 and the distribution of product of the coefficients method.^100,101 Mediation proportions were limited to 0–100%, with values below 0% set to 0% and those above 100% capped, and only combinations with mediation proportions above 0% were included.

Exposure and covariate measurements in observational study

Data on various female reproductive traits, such as AFB, ALB, AFS, AAM, NLB, age at menopause, preterm delivery status, and infertility were collected via self-reports. AFB was assessed by asking, “How old were you at the time of your first live birth?” Responses indicating 14 years or younger were recorded as 14. ALB was evaluated by asking, “How old were you at the time of your last live birth?” Responses of 14 years or younger were recorded as 14, while those 45 years or older were recorded as 45. AFS was determined through the question, “How old were you when you had sex for the first time?” Responses of 12 years or younger were recorded as 12. Age at menarche was recorded by asking, “What was your age when your first menstrual period occurred?” Responses of 20 years or older were recorded as 20, and individuals who had not yet begun menstruation or responded with 0 were excluded. Age at menopause was determined by asking, “About how old were you when you had your last menstrual period?” Responses of 19 years or younger were recorded as 19, and those 60 years or older were recorded as 60. Infertility was evaluated by asking participants if they had ever experienced a year or more of unsuccessful attempts to conceive and if they had sought medical assistance for difficulties in becoming pregnant. Women who responded positively to either question were classified as having a history of infertility. Preterm delivery status was available only for women who had delivered low-birth-weight infants and was assessed by asking, “How many of these babies were born preterm? A preterm delivery is one that occurs at 36 weeks or earlier in pregnancy.” If an individual had one or more preterm deliveries, they were categorized as having a history of preterm birth. The number of live births was determined by asking, “How many of your deliveries resulted in a live birth?” Responses of 11 or more were recorded as 11. For gestational diabetes, endometriosis, and uterine leiomyoma, if a doctor informed the patient of the condition, it was recorded as “yes”. The number of pregnancy losses (PLN) was calculated by subtracting the count of live birth deliveries from the total pregnancies initiated. ANM was defined for participants as the age at menopause, specifically for women aged 40 to 60 years. Women who had hysterectomy or bilateral oophorectomy were excluded.¹⁰² The number of years ovulating was calculated using the following formula: Years ovulating = years menstruating - years on pill - 0.77 × number of live births - 0.25 × number of pregnancy losses. In this formula, years menstruating is calculated by subtracting menarche age from menopause age, and years on pill refers to the number of years using birth control pills, with a value of 0 if never used. For women with no pregnancies, both the number of live births and the number of pregnancy losses are recorded as 0. The factor 0.77 accounts for the 40 weeks of pregnancy divided by the 52 weeks in a year, while the factor 0.25 reflects the average duration of an unsuccessful pregnancy, calculated as 3 months out of 12.¹⁰³

Hormone levels were assessed using standard laboratory methods. Total testosterone concentrations were assessed via isotope dilution liquid chromatography-tandem mass spectrometry (ID-LC-MS/MS). SHBG levels were quantified through an immuno-antibody reaction, followed by chemiluminescent detection. This process required two incubation phases and the application of a magnetic field. Bioavailable testosterone levels were calculated using Vermeulen et al.’s formulas,¹⁰⁴ which are based on serum testosterone and albumin concentrations. These formulas can be accessed on the ISSAM website (https://www.issam.ch/freetesto.html). Bioavailable testosterone includes all testosterone not bound to SHBG, making it accessible for biological activity. Serum FSH and LH levels were measured using microparticle enzyme immunoassay technology.

Our study included various covariates, with age defined in years. Race and ethnicity were determined based on self-reported data. Socioeconomic status was evaluated through the poverty-to-income ratio (PIR), categorized into four groups: below 1.3, from 1.3 to 3.5, above 3.5 and no record. Educational attainment was grouped into three levels: less than high school, high school graduate, and more than high school. Marital status was classified as divorced/separated/widowed, married/living with a partner, never married or no recorded status. Body mass index (BMI) determined by dividing weight (kg) by height (m²) was grouped into four categories: underweight (BMI <18.5) and normal weight (BMI 18.5–25), overweight (BMI 25–30), and obese (BMI ≥30). Smoking status was divided into three groups: never (under 100 cigarettes throughout one’s life), former (over 100 cigarettes smoked in a lifetime, but not a current smoker), and now (over 100 cigarettes and currently smoking, either occasionally or daily). Alcohol intake was classified into the following categories: never (individuals who had consumed fewer than 12 drinks throughout their lifetime), former (individuals who drank at least 12 drinks in a year but had abstained from alcohol in the last year), mild (those who consumed one drink per day for females), moderate (those who had consumed up to two drinks per days for females or had two to four binge drinks), and heavy (those who had consumed three or more drinks per days for females or had five or more binge drinks).¹⁰⁵

Medical history, including hypertension, diabetes, cardiovascular disease, and cancer, was obtained from medical diagnoses provided by doctors or other healthcare professionals. Information regarding the removal of the ovary (either unilateral or bilateral), hysterectomy, use of female hormones, and pregnancy history was self-reported during interviews. Menopausal status was determined by participants’ reported reasons for irregular periods or by their age, with those reporting menopause or being older than the menopausal age classified as postmenopausal. The live birth category was determined based on the count of births leading to live infants and was divided into three groups: no live births, fewer than three live births, and three or more live births.

Outcome measurements in observational study

Due to the limitations of the NHANES dataset, which includes participants only up to age 85, the 90th and 99th percentile survival ages required for longevity GWAS could not be estimated. These estimates typically require data from individuals aged 85 to over 100 years, calculated from life tables like those in the Human Mortality Database.¹⁰⁶ As a result, mortality data were used as a proxy for longevity studies. Mortality information was gathered by connecting the cohort database with the National Death Index, which included deaths up to December 31, 2019. Deaths from all causes were classified as all-cause mortality.

Biological age was assessed using two approaches. DNAm data were obtained from biospecimens of individuals aged 50 years and above during the 1999–2000 and 2001–2002 cycles and processed using Illumina EPIC BeadChip arrays. After preprocessing and normalization, epigenetic biomarkers derived from DNAm data were utilized to predict biological age. These encompass both first- and second-generation epigenetic clocks consistent with MR to capture various aspects of biological aging. The first-generation clocks include DNAmHannumAge, IEAA derived from DNAmHorvathAge, and DNAmHorvathAge without blood cell correction, as well as Horvath’s DNAm-predicted chronological age in skin and blood-derived tissues (DNAmSkinBloodAge)¹⁰⁷ for validation. The second-generation clocks include DNAmPhenoAge, DNAmGrimAge, and an updated version, DNAmGrimAge2, which improves mortality prediction in whole blood.¹⁰⁸ Because of the absence of DNAm-based biological age acceleration data for infertility, gestational diabetes, and androgen-related markers from 1999 to 2002, we employed an alternative approach to assess biological aging. The second method relies on blood chemistry and body measurement data. This involved two validated algorithms: the Klemera-Doubal method (BioKDMAge)¹⁰⁹ and the phenotypic age calculation (BioPhenoAge)¹¹⁰ using clinical biomarkers, both computed with the R package BioAge.¹¹¹ BioKDMAge was calculated using eight biomarkers: HbA1C, albumin, alkaline phosphatase, creatinine, total cholesterol, systolic blood pressure, CRP, and blood urea nitrogen. BioPhenoAge was derived from nine blood chemistry biomarkers: white blood cell count, lymphocyte percentage, red cell distribution width, albumin, CRP, glucose, mean cell volume, alkaline phosphatase, and creatinine.

We computed age acceleration (AgeAccel) for each biological age as the residual from a linear regression of biological age on chronological age. For instance, DNAmGrimAgeAccel indicates if an individual is biologically older (positive value) or younger (negative value) than expected for their chronological age. IEAA, derived from the Horvath clock, represents cell-intrinsic aging independent of blood cell type composition, which was assessed using the Houseman algorithm to estimate the proportions of white blood cells, including CD8⁺ T-cells, CD4⁺ T-cells, natural killer cells, B-lymphocytes, monocytes, and neutrophils.¹¹² IEAA was calculated by regressing DNAmHorvathAge on chronological age, adjusted for these blood cell proportions.¹¹³

Statistical analyses in observational study

All data analyses complied with the NHANES Analytic Guidelines, utilizing procedures to account for NHANES’s intricate sampling design, with sample weights, clustering, and stratification integrated into every analysis.¹¹⁴ The baseline characteristics of participants were reported as weighted means with standard errors for continuous variables and as numbers with weighted proportions for categorical variables, with all values adjusted for sampling weights. To evaluate the association between female reproductive traits and all-cause mortality, weighted multivariable Cox proportional hazards regression models were employed to estimate hazard ratios (HRs) with 95% CIs. Person-years for each participant were were determined from the enrollment date until either the date of death or the end of follow-up (December 31st, 2019), whichever occurred first. The association between female reproductive traits and biological aging was analyzed using weighted linear regression models.

Three multivariate models were constructed to adjust for potential confounders. In Model 1, adjustments were made for age (continuous, years), race and ethnicity, PIR, educational level, and marital status. Model 2 included additional adjustments for alcohol intake, BMI category, smoking status, and history of hypertension, diabetes, cardiovascular disease, and cancer. Model 3 further adjusted for history of ovary removal, pregnancy history, menopause status, female hormone use, live birth category, and history of hysterectomy. Following recommendations by Nancy Krieger et al., models using AgeAccel as the outcome were still adjusted for chronological age.¹¹⁵ For traits such as AFB, ALB, preterm status, NLB, and PLN, the analysis population consisted of women who had given birth, negating the need for pregnancy history adjustments. While other analyses adjusted for menopause status, in our analyses of age at menopause, years of ovulation, and ANM within a population where all participants had already undergone menopause, such adjustments were unnecessary. When analyzing NLB, the live birth category was excluded as a covariate, and for ANM, adjustments for ovary removal or hysterectomy history were not included. Multicollinearity in Cox Model 3 was mitigated by excluding reproduction-related covariates with high VIFs. In the SHBG and BAT models, pregnancy history and live birth category were removed, retaining only female hormone use in the non-linear model. For the LH and FSH models, pregnancy history, live birth category, and menopause status were excluded, leaving no reproductive covariates in the non-linear models to prevent multicollinearity. Non-linear associations between female reproductive traits, all-cause mortality, and biological aging were assessed using restricted cubic spline curves. The number of knots for spline curves was selected between three and seven to optimize model fit without overfitting. Knots were selected based on the lowest Akaike Information Criterion (AIC) value when non-linearity was significant (P-nonlinear <0.05), with preference given to models with fewer knots if the AIC values differed by less than two.¹¹⁶ Given the relatively narrow age range of natural menopause, with the majority (70%) of women experiencing menopause between 47 and 55 years of age, we narrowed the knot selection range for ANM to 3–5. The final number of knots used in each group is presented in Table S24.

Quantification and statistical analysis

All MR analyses were conducted using the R software (version 4.2.2; Vienna, Austria), with the following packages: TwoSampleMR (v0.6.3), MVMR (v0.4), MRPRESSO (v1.0), MendelianRandomization (v0.9.0), and RMediation (v1.2.2). In the UVMR analysis evaluating the casual relationship between female reproductive traits and aging, a Bonferroni correction was applied to account for multiple testing. Statistical significance was defined at three levels: nominal (p < 0.05), trait-adjusted Bonferroni (p < 0.05/37 reproductive traits = 1.35 × 10⁻³), and phenotype-adjusted Bonferroni (p < 0.05/37/6 aging phenotypes = 2.25 × 10⁻⁴). MR results were expressed as odds ratios (ORs), β coefficients, or proportions, each accompanied by their respective 95% confidence intervals (CIs). All observational statistical analyses were conducted using R (version 4.2.2; Vienna, Austria), with statistical significance determined by a two-sided p value of less than 0.05.

Published: December 16, 2025