Do Epigenetic Clocks Provide Explanations for Sex Differences in Life Span? A Cross-Sectional Twin Study

Abstract Background The sex gap in life expectancy has been narrowing in Finland over the past 4–5 decades; however, on average, women still live longer than men. Epigenetic clocks are markers for biological aging which predict life span. In this study, we examined the mediating role of lifestyle factors on the association between sex and biological aging in younger and older adults. Methods Our sample consists of younger and older twins (21‒42 years, n = 1 477; 50‒76 years, n = 763) including 151 complete younger opposite-sex twin pairs (21‒30 years). Blood-based DNA methylation was used to compute epigenetic age acceleration by 4 epigenetic clocks as a measure of biological aging. Path modeling was used to study whether the association between sex and biological aging is mediated through lifestyle-related factors, that is, education, body mass index, smoking, alcohol use, and physical activity. Results In comparison to women, men were biologically older and, in general, they had unhealthier life habits. The effect of sex on biological aging was partly mediated by body mass index and, in older twins, by smoking. Sex was directly associated with biological aging and the association was stronger in older twins. Conclusions Previously reported sex differences in life span are also evident in biological aging. Declining smoking prevalence among men is a plausible explanation for the narrowing of the difference in life expectancy between the sexes. Data generated by the epigenetic clocks may help in estimating the effects of lifestyle and environmental factors on aging and in predicting aging in future generations.

When the lifestyle factors (education, BMI, smoking, alcohol use, and leisure index) were controlled for each other in the multiple mediator models, higher BMI was still associated with higher AAHorvath, AAPheno, and AAGrim (B=0.11-0.15) ( Figure 4). Smoking was associated with accelerated AAHannum and AAPheno only in the older twins (B=0.01 and smoking × age: B=0.13 and B=0.06 and smoking × age: B=0.13, respectively), whereas smoking was associated with accelerated AAGrim also in the younger twins, and the association was stronger in the older twins (B=0.47, smoking × age: B=0.14). Greater alcohol use was associated with accelerated AAPheno only in the older twins (B=0.01; alcohol × age: B=0.06). Education and leisure index were not associated with AA after controlling for other lifestyle-related factors. The direct effect of male sex was still observed on AAHorvath, AAHannum and AAGrim, and the association was stronger in the older twins (B=0.08 to 0.13, sex × age: B=0.07 to 0.14). Instead, male sex was associated with slower AAPheno in the younger twins, but the association turned positive in the older twins (B=-0.11, sex × age: B=0.23).

Sensitivity analyses
To confirm that observed sex differences in mediator variables and consequently observed indirect effects are not due to the differences in the age distribution between sexes, we re-analyzed the data using polynomial models of age. The interpretation of the results considering sex differences in lifestyle-related factors was very similar to those obtained in the main analysis when a dichotomous variable of age was used (Supplementary Table S3 and Figure S4). Association between age and education followed a quadratic curve. There was a difference in a linear term between sexes (Sex × Age) showing that men were better educated than women in older age. Association between age and BMI followed a quadratic curve, as well. Overall, men had higher level of BMI, and no significant interactions of sex and age were observed. Association between age and (latent) smoking was linear. Men had higher level of smoking and the difference appeared to increase with age, but interaction (Sex × Age) did not reach significance (p=0.066). Association between age and alcohol use followed a cubic curve. There was a difference in both a linear term and quadratic term between sexes (Sex × Age and Sex × Age 2 ). Overall, men consumed more alcohol than women ( Figure S4). The difference was largest at younger age but widened again after 50 years of age. The associations between age and both sport and leisure index were linear. Men had lower level of leisure index, but contrary to our analysis using dichotomous age, there were no significant sex differences in sport index. Association between age and work index followed a quadratic curve and there were no significant sex differences in work index.
The estimation results of the single mediator models using polynomial function of age are presented in Supplementary  Table S4. The indirect effects by age were visually inspected using loop plots (Supplementary Figures S5-S17). Interpretation was very similar to the results of the main analysis using a dichotomous variable of age, but there were few exceptions. Education mediated the sex difference only in younger age and only when AAGrim was used to assess epigenetic aging (Supplementary Figure S5). Body mass index did not mediate the sex difference (Supplementary Figure S6). Alcohol use mediated the association only when AAPheno was used and only in older twins (Supplementary Figure S8).
The estimation results of the multiple mediator models using polynomial function of age are presented in Supplementary Figure S12 and the indirect effects by age in Supplementary Figures S13-S17. These results were very similar to the main results, as well. In line with the main results, also BMI appeared to mediate the sex difference in AAHorvath, AAPheno and AAGrim (Supplementary Figure S14).. This inconsistency between the results of the single and the multiple mediator models of sensitivity analysis is probably due to the fact that insignificant interaction terms were dropped from the multiple mediator models. The only difference to the main analysis was that education did not mediate the association when AAGrim was used to assess epigenetic aging (Supplementary Figure S13).

Mediation models in the opposite-sex twin pairs
In the opposite-sex twin pairs, higher BMI was associated with accelerated AAHorvath (B=0.24) (Supplementary Table  S5). Smoking was associated with accelerated AAGrim (B=0.36) and higher sport index with slower AAHorvath (B=-0.14). In all the mediation models, direct association of sex with AAHannum and AAGrim was observed (B=0.22-0.29). When the lifestyle factors (education, BMI, smoking, alcohol use, and leisure index) were controlled for each other in the multiple mediator models, higher BMI was still associated with accelerated AAHorvath (B=0.26) and smoking with accelerated AAGrim (B=0.35) ( Figure 5). Surprisingly, higher leisure index was associated with accelerated AAHannum (B=0.18). A positive direct effect of male sex on higher AAHannum and AAGrim was observed (B=0.29 and B=0.19, respectively).

Sex differences in DNAm-based plasma proteins and smoking pack-years
Information on sex is utilized in the estimation of epigenetic age by GrimAge estimator (3). Therefore, the observed sex difference in AAGrim may reflect the estimated sex difference in mortality and not only the differences in DNAm. Also, in DNAm-based surrogates included in the GrimAge estimator the sex difference is in-built, indicating differences between men and women in the actual levels of plasma proteins and smoking pack-years. To further understand the sex differences in biological aging, we also studied the differences in age-adjusted DNAm-based plasma proteins and smoking pack-years in all twins and in the opposite-sex twin pairs. The variables were standardized before the analysis.
For DNAm-based ADM and B2M, a linear model was sufficient (Supplementary Table S6). Men had a lower level of DNAm ADM and B2M in young adulthood, but the difference narrowed or disappeared with age (Supplementary Figure S18, A-B). For DNAm GDF15, a quadratic model was required, and there was a sex difference in the second order term (Supplementary Table S6). Among women the association followed U-shaped pattern, whereas among men the level of DNAm GDF15 decreased linearly with age (Supplementary Figure S18, C). The men had slightly lower level of DNAm GDF15 in younger age. The difference disappeared after 30 years of age but widened again after 50 years of age. Association between age and DNAm cystatin C followed a quadratic curve, and there was a sex difference in the linear term (Supplementary Table S6). The sex difference in this surrogate increased rather steeply from midlife onwards and the men had a higher level especially in older age (Supplementary Figure S18, D). For DNAm leptin the difference in a third order term between the sexes was significant in a cubic model (Supplementary Table S6). Overall, the men had a lower level of DNAm leptin (Supplementary Figure S18, E). In younger age, the sex difference was constant across ages but after 50 years of age the difference slightly narrowed. For DNAm PAI-1 a cubic model was required (Supplementary Table S6). Overall, the men had higher level of DNAm PAI-1, and the sex difference increased rather steeply with age (Supplementary Figure S18, F). For DNAm TIMP-1 and DNAm packyrs a linear model was sufficient (Supplementary Table S6). The sex difference in DNAm TIMP-1 increased with age and the men had a higher level of DNAm TIMP-1 especially in older age (Supplementary Figure S18, G). The men had also a higher level of DNAm-based smoking pack-years, and the difference did not depend on age (Supplementary Figure S18, H).
In the opposite-sex twin pairs, the men had significantly lower levels of DNAm ADM, B2M, and leptin, whereas higher levels of DNAm PAI-1 and DNAm-based smoking pack-years (Supplementary Figure S19). Figure S2. Correlation coefficients among epigenetic age acceleration (AA) measures.

Supplementary
Supplementary Table S1. The association between age and epigenetic age acceleration (AA) modelled as a third to first order polynomial function of age (n = 2240).  Notes: ADM, adrenomedullin; B2M, beta-2 microglobulin; GDF15, growth differentiation factor 15; PAI-1, plasminogen activation inhibitor 1; TIMP-1, tissue inhibitor metalloproteinase 1; packyrs, pack-years. B, standardized (STDYX) regression coefficient; SE, standard error. a The model was controlled for zygosity. b SEs were corrected for nested sampling. c The higher order interactions (Sex × Age 2 , Sex × Age 3 ) were freed only when necessary (modification index > 4) d The model was not fitted, because a higher order polynomial model was needed for the measure. e The model did not include the corresponding polynomial term.