Second generation DNA methylation age predicts cognitive change in midlife: the moderating role of childhood socioeconomic status

Predicting cognitive outcomes from DNAmAge

Individuals with accelerated DNAmAge have an increased risk of cancer, stroke, heart disease, and neurodegenerative disease [4–6]. While cognitive functioning is a crucial facet of both healthy and pathological aging, the link between DNAmAge and cognitive functioning remains unclear, partly due to inconsistencies in the predictive accuracy of various DNAmAge algorithms [7]. First generation DNAmAge measures, such as Horvath, Horvath Skin and Blood, and Hannum DNAmAge, were designed primarily to estimate chronological age with methylation data from cytosine-phosphate-guanine (CpG) sites sensitive to time-dependent changes [1, 8, 9]. While two studies found that Horvath DNAmAge was associated with cognitive decline [6, 10], several others have found no statistically significant associations between first generation DNAmAge measures and cognitive performance or decline [11–14].

In contrast, second generation DNAmAge measures, including PhenoAge and GrimAge, are trained to predict age-related outcomes by integrating DNAm signatures tied to clinical biomarkers, health behaviors (e.g., smoking), and demographic variables [15, 16]. These algorithms aim to capture phenotypic aging processes beyond chronological age, such as morbidity and mortality risk. Accelerated PhenoAge and GrimAge are consistently associated with lower cognitive performance in middle and older adulthood [11, 12, 17, 18]. However, findings on cognitive decline are mixed. Neither Hillary et al. [18] nor Maddock et al. [12] reported statistically significant results, yet Reed and colleagues [13] found that those with cognitive decline in midlife showed more accelerated GrimAge scores than those without decline. These findings suggest that second generation measures, developed to more accurately estimate phenotypic biological age, may better index age-related changes in cognitive ability.

Early-life exposures and pace of DNAmAge

Disadvantaged socioeconomic status (SES) in childhood is an established risk factor for lower cognitive ability across the lifespan and late-life cognitive decline [19]. Recent research has also begun exploring links between childhood adversity and later-life DNA methylation age (DNAmAge). In retrospective accounts of adversity (e.g., poverty, death of parent, alcohol/drug use in family, abuse), only poverty was related to GrimAge acceleration [20]. Childhood SES but not adult SES predicted accelerated DNAmAge in midlife using the Horvath and Hannum clocks, suggesting the methylome, or DNA methylation modifications across the genome, may be particularly vulnerable to external stressors in childhood [21, 22]. However, findings remain inconsistent: McCrory et al. [23] found no association between childhood SES and Horvath, Hannum, or PhenoAge DNAmAge in adults over 50, and unexpectedly, Faul et al. [11] found that low childhood SES was associated with slower aging on the Horvath and Hannum clocks later in life. It is not yet known whether SES moderates the effects of DNAmAge on cognitive outcomes.

Genetics, epigenetics, and classical twin designs

Associations between early life exposures, mid- to late-life outcomes, and DNAmAge have primarily been correlational. Two studies decomposed the variance of DNAmAge into additive genetic, shared environmental, and nonshared environmental components using a traditional twin model [24, 25], while two others used pedigree designs to estimate heritability [16, 26]. Our group has shown how to use a longitudinal twin design to go beyond estimating heritability while sharpening causal inference between a predictor and outcome [27]. Only two studies have examined the relationship between DNAmAge and cognitive ability while controlling for between-family confounds [10, 14]. In the Middle-Aged Danish Twin Study, no relationship was observed between the Horvath and Hannum clocks and cognitive change over 10 years in midlife [14]. Vaccarino and colleagues [10] found no associations between Horvath, Hannum, GrimAge, or PhenoAge clocks and baseline cognition in middle-aged male twins. However, there was a statistically significant within-twin pair effect of Horvath DNAmAge on cognitive decline 11 years later. The reason why only Horvath’s original measure predicted cognitive ability is unclear, as initial validation studies of PhenoAge and GrimAge reported associations with cognitive outcomes [15, 16].

Using the Louisville Twin Study (LTS), we expanded on this limited amount of research examining DNAmAge and cognitive change and sought to clarify how early life environmental characteristics may influence midlife cognitive outcomes. The LTS, initiated in 1957, includes intensive, prospective data on twins’ development from infancy through adolescence and recent re-evaluations in midlife on their cognitive, physical, and psychosocial functioning as well as epigenetic aging [28]. Parental data collected during the twins’ childhood further enabled us to examine how early life exposures, here rearing SES, relate to twins’ midlife epigenetic and cognitive aging.

The current study used a twin design to investigate causal effects of DNAmAge on change in IQ (ΔIQ) between childhood and midlife, while testing childhood SES as a moderator of the relationship. While observational designs cannot reproduce the causal inference possible with random assignment, twins offer a quasi-experimental method for studying causal processes while controlling for family-level genetic and environmental confounds. If a monozygotic (MZ) twin has a higher DNAmAge and also a greater ΔIQ than their co-twin, the association cannot be due to genetic differences or early environmental exposures, because identical twins reared together are matched for these factors. For this reason, we refer to associations that survive twin-based controls as “quasi-causal” [27]. To address variability in DNAmAge algorithms, we conducted an exploratory factor analysis of five widely used measures. We expected second generation DNAmAge measures to be sensitive to ΔIQ in midlife, as they were developed to capture phenotypic aging processes beyond chronological age. Finally, given recent findings linking early life stressors with midlife DNAmAge, we hypothesized that the relationship between midlife DNAmAge and ΔIQ would depend on differences in childhood SES.

Descriptive statistics

The sample includes 287 individual twins, comprising 60 monozygotic and 41 dizygotic complete pairs and 85 single twins. All participated in the childhood and midlife phases of the LTS (Supplementary Figure 1). Twins were assessed in person using three versions of the Wechsler Intelligence Scale for Children (WISC) and the Wechsler Adult Intelligence Scale (WAIS-IV). We used Wechsler Full-Scale Intelligence Quotient (IQ) scores as an overall estimate of an individual’s level of global cognitive ability. The average age for childhood IQ was 14.42 (1.93) years, and the average age for adult IQ was 51.87 (7.03). Mean childhood IQ was 102.7 (13.4), and mean adult IQ was 105.4 (13.8) (Table 1). DNAmAge was estimated using principal components (PC) based epigenetic clock methodology which has shown significantly improved reliability and stability over time as compared with the traditional Horvath, Horvath Skin and Blood, Hannum, PhenoAge, and GrimAge algorithms [29]. After regressing DNAmAge values on chronological age and adjusting for cell type, Horvath, Horvath Skin and Blood, Hannum, PhenoAge, and GrimAge had means close to zero (Table 1). MZ twin correlations (0.60-0.74) were approximately double those for DZ twins (0.28-0.37) across most measures, except PhenoAge, where MZ and DZ correlations were more similar (0.56 vs. 0.47). Pairwise clock correlations are in Supplementary Table 1.

Results from standardized univariate twin models decomposing the variance of individual DNAmAge measures showed little to no shared environmental variance, C, in Horvath, Horvath Skin and Blood, Hannum, and GrimAge clocks. Because C was estimated at a negative value in four of five ACE models initially, C was set to zero, and the constrained AE models are presented (Supplementary Table 2). The five DNAmAge measures were substantially heritable, with the additive genetic effects, A, ranging from 57% to 73%, and non-shared environment, E, accounting for 27% to 43% of the variance in DNAmAge (Supplementary Table 2).

Exploratory factor analysis of DNAmAge measures

We conducted an exploratory factor analysis (EFA) using the five DNAmAge measures residualized for age and cell count. Based on Eigenvalues and fit indices, we selected a two-factor model (RMSEA = 0.08; 90% CI: [0.00, 0.20]) (Supplementary Figure 2). Results from a likelihood ratio test comparing a two-factor model to a one-factor model suggested that the two-factor model fit substantially better (χ² =70.07, df=4, p < .001). First generation DNAmAge measures (Horvath, Horvath Skin and Blood, and Hannum) loaded on one factor, while second generation measures (PhenoAge and GrimAge) loaded on a second factor. We refer to these factors as “Gen 1” and “Gen 2”, respectively. Factor loadings (p < .05) are displayed in Supplementary Table 3. The factors were correlated r = 0.57. Composite scores were generated for each participant based on the EFA. Gen 1 is the sum of standardized values for Horvath, Horvath Skin and Blood, and Hannum DNAmAge (M=0.00, SD= 2.85). Gen 2 is the sum of standardized values for GrimAge and PhenoAge DNAmAge (M=0.00, SD= 1.78). Twin pair correlations for Gen 1 and Gen 2 are in Table 1.

Univariate twin models of Gen 1 and Gen 2 DNAmAge

We decomposed the variances of Gen 1 and Gen 2 into A, C, and E components using the classical twin model. Both Gen 1 and Gen 2 were moderately heritable, with A accounting for 50% of the variance in Gen 1 and 60% of the variance in Gen 2. The variance attributable to the non-shared environment was 42% in Gen 1 and 31% in Gen 2, while shared environment accounted for less than 1% of the variance in both. Standardized estimates are in Supplementary Table 2.

Phenotypic and “quasi-causal” association models

Following the twin modeling framework outlined by Turkheimer and Harden [27], we first estimated a “phenotypic association” model with no control for differential regressions between and within twin pairs. We then fit a “quasi-causal” regression model, using the ACE decomposition of the predictor variable to control for family-level genetic and environmental confounds in the individual-level regressions. This sequence (phenotypic association model and quasi-causal regression model) was estimated for Gen 1 then repeated for Gen 2. Figure 1 illustrates the model: adult IQ was regressed on childhood IQ, DNAmAge, and the A and C components of each; DNAmAge was regressed on childhood IQ and its A and C components; SES (a family-level variable) was included as a linear covariate for DNAmAge and IQ variables and as an interaction term with DNAmAge. By including childhood IQ in the model, the conditional effects of SES and DNAmAge can be interpreted as predicting change in IQ from childhood to midlife (ΔIQ). The phenotypic regression coefficient (b_P3) is an estimate of the relationship between the DNAmAge and ΔIQ with between-family genetic and environmental confounds controlled, i.e., within a pair of identical twins reared together.

Figure 1. Path diagram depicting the association between DNAmAge and change in IQ, modified by childhood socioeconomic status (SES). Variances of ACE components are estimated, with paths to observed variables set to 1.0. All regression coefficients are unstandardized. b_A1 and b_C1 are regression coefficients from the A and C components of Childhood IQ to Adult IQ. b_P1 is the phenotypic regression conditional on b_A1 and b_C1. b_A2 and b_C2 are regression coefficients from the A and C components of Childhood IQ to DNAmAge. b_P2 is the phenotypic regression conditional on b_A2 and b_C2. b_A3 and b_C3 are regression coefficients from the A and C components of DNAmAge to Adult IQ. b_P3 is the phenotypic regression conditional on b_A3 and b_C3. SES is a covariate for DNAmAge and the IQ variables, and a moderator of b_P3.

Predictors of DNAmAge

Phenotypic models. In the phenotypic association model, there were no statistically significant associations with Gen 1 DNAmAge (Supplementary Table 4). However, accelerated Gen 2 DNAmAge was associated with lower childhood SES (b=-0.41, SE=0.13, p<.05) (Table 2).

Quasi-causal models. Next, we used the quasi-causal regression model to estimate the effects of childhood IQ and SES on DNAmAge, using the ACE decomposition of childhood IQ to control for genetic and environmental family-level confounds. There were similarly no statistically significant predictors of Gen 1 DNAmAge. In the Gen 2 quasi-causal analyses, childhood SES again predicted Gen 2 DNAmAge (b=-0.67, SE=0.19, p<.001), while childhood IQ did not predict Gen 2 DNAmAge (Table 2). Genetic and environmental confounds were not significantly different from zero.

Predictors of midlife cognitive functioning

Phenotypic models. In the phenotypic association models for both Gen 1 and Gen 2, childhood IQ statistically predicted adult IQ. The strong relationship between childhood IQ and adult IQ (r = 0.77) is also illustrated in Supplementary Figure 3, where adult IQ is regressed on childhood IQ without controlling for the differential regressions between and within twin pairs. Similarly, SES was positively associated with childhood IQ (b=0.56, SE=0.09, p <.001) in both phenotypic association models. Higher childhood SES predicted higher adult IQ in the Gen 1 model (b=0.16, SE=0.07, p<.05), though this association did not meet statistical significance in the Gen 2 model (Supplementary Table 4). Gen 1 and ΔIQ were not associated with each other (Supplementary Table 4). We did find a statistically significant phenotypic association between Gen 2 and ΔIQ (b_P3= -0.10, SE = 0 .03, p <.05). In both Gen 1 and Gen 2 phenotypic association models, we then tested the interaction between DNAmAge and SES (b_int) in the prediction of ΔIQ but found that the interaction was not statistically significant.

Strengths and limitations

Unlike previous studies that relied on retrospective reports of childhood financial hardship or parental occupation, ours is the first to use a prospective rating of childhood SES, which may have reduced recall bias and improved reliability. However, our childhood SES measure was a relatively crude indicator of parental occupation, and incorporating additional socioeconomic factors could strengthen future research. We did not test adult SES in this study, given that previous work had suggested childhood SES, but not adult SES, was predictive of DNAmAge [21, 22]. While our sample covered a broad socioeconomic range, it was predominantly white and recruited from one region of the United States, limiting generalizability. Future studies should examine the relationship between second generation DNAmAge measures and patterns of cognitive ability in larger and more diverse samples. Finally, we used IQ score as a proxy for global cognitive ability and future work should examine relationships between DNAmAge and specific cognitive domains including memory and executive functioning.

Participants

Participants were 287 individual twins from the midlife phase of the Louisville Twin Study (LTS). The LTS began as a study of childhood development in multiple birth pairs. Twins born between 1950-1997 in the Louisville metropolitan area were followed and assessed on cognitive ability, physical development, and temperament between 3 months and 15 years of age. Parents of twins provided data on demographics, parenting factors, and home environment. Zygosity was determined through blood typing during childhood. The sample contains same-sex monozygotic (MZ) and dizygotic (DZ) twin pairs, and opposite-sex DZ pairs. The LTS closed down in 2000 after 1,770 individual twins had been recruited [35]. In 2018, follow-up of LTS twins aged 30-65 was initiated [28], and LTS reopened in 2019 to begin the first midlife data collection of the 1,770 twins. Participants provided a 50cc whole blood sample for genotyping and DNA methylation assay. Genotyping confirmed zygosity for any unknown twin pairs. The present sample includes 60 MZ pairs, 41 DZ pairs, and 85 individual twins who completed a midlife study visit between 2019-2023. Singletons were included in the analysis and contributed to between-pair analyses but not to within-pair regressions. Supplementary Figure 1 illustrates the participant flow, attrition, and retention across study phases.

Socioeconomic status

Childhood socioeconomic status (SES) was measured using the Duncan socioeconomic index which assigns scores of 0-100 based on the occupation of the head of the household [36]. Occupational prestige was based on 1950 US Census income data and education level associated with the occupation. The childhood SES of our sample (M=47.3, SD = 25.7) was consistent with the overall LTS distribution (M=46.9, SD= 26.9), evenly distributed, and representative of the Louisville, Kentucky area [35].

Smoking

During the midlife assessment, participants self-reported whether they had smoked tobacco in the last three years. We coded responses at 1 (smoked in the last three years and/or current smoker) and 0 (has not smoked in the last three years).

Cognitive ability measures

LTS participants were assessed using validated Wechsler batteries of cognitive functioning. Children and adolescents were administered one of three forms of the Wechsler Intelligence Scale for Children (WISC, WISC-R, WISC-III) [37–39]. Adults in the midlife phase were assessed using the Wechsler Adult Intelligence Scale IV (WAIS-IV) [40]. Test scores are age-standardized based on a standardization sample that is roughly representative of the United States. Wechsler IQ scores have a mean of 100 and a standard deviation of 15. Children were typically administered cognitive assessments at ages 7, 8, 9, 12, and 15, though not all twins were assessed at each timepoint; this study used the most recently available WISC assessment for each participant, with over 80% collected at age 15 (see Table 1 for more detailed descriptive statistics).

DNAmAge measures

Genomic DNA was extracted from whole-blood samples that were collected at the University of Louisville and Norton Healthcare medical campuses. Venous blood was collected in ethylenediaminetetraacetic acid (EDTA) tubes and shipped to the Norris Comprehensive Cancer Institute at the University of Southern California Keck School of Medicine. DNA was extracted with promega Maxwell 16 LEV blood DNA kits, and then treated with bisulfite reagents at the USC Molecular Genomics Core following manufacturer protocol. Methylation was assayed with the Illumina Infinium Human MethylationEPIC BeadChip (Illumina, San Diego, CA, USA) at the USC Molecular Genomics Core at the Norris Cancer Institute at the Keck School of Medicine.

Quality control was conducted in R 4.3.0 [41] using the minfi package [42] to identify aberrant samples and CpG sites, remove cross-reactive probes, conduct background correction, and adjust for batch effects prior to estimating DNAmAge variables. Background correction was performed using the normal-exponential out-of-band (noob) method [43]. The ComBat method was used to adjust for laboratory batches [44].

We estimated blood cell composition using the Houseman method [45] and obtained percentages of CD8+ T cells, CD4+ T cells, natural killer cells (NK), B cells, monocytes, and granulocytes. Cell proportions are highly correlated, so in the current analysis we included the first and second principal components from a principal components (PC) analysis (collectively, they explained 86.12% of the variance) in all analyses.

As replicate blood samples often have different methylation values due to noise in the CpG sites, batch effects, and sample preparation, Higgins-Chen et al. [29] developed a principal components-based method to bolster reliability in DNAmAge algorithms by training them based on PC analysis, extracting the covariance between multicollinear CpGs, including age-related covariance. We estimated DNAmAge using five PC-trained algorithms, Horvath, Horvath Skin and Blood, Hannum, PhenoAge, and GrimAge [1, 8, 9, 15, 16].

Data analysis

Descriptive statistics for the sample’s demographic characteristics, survey data, and cognitive ability data were computed in R version 4.3.2 [41]. Methylation data were cleaned and prepared in R. DNAmAge was calculated as the residual difference in PC-trained values predicted from chronological age in a bivariate linear regression model. DNAmAge values were adjusted for cell composition. Age and cell-adjusted variables were then winsorized such that values more than two standard deviations from the mean DNAmAge score were replaced with a value of two standard deviations above or below the mean. Pairwise correlations for the five DNAmAge measures were then estimated using maximum likelihood in Mplus Version 8.8 [46].

We used the classical twin model to decompose the phenotypic variance of DNAmAge into additive genetic (A), shared environmental (C), and nonshared environmental (E) components. Given that MZ twins share 100% of their genetic makeup and DZ twin pairs share half, on average, the A variance is correlated 1.0 within MZ twin pairs and 0.5 within DZ twin pairs. In both MZ and DZ twin pairs, the C component includes experiences that make siblings raised together more similar, and is correlated 1.0. Nonshared environmental components are unique to individuals and by definition uncorrelated within twin pairs. We estimated a univariate ACE model for all five DNAmAge measures individually using Mplus. Because C values were estimated as negative, we selected to use the constrained AE models, with C set to zero. AE estimates with 95% confidence intervals calculated using bias corrected bootstrap standard errors are presented in Supplementary Table 2.

Exploratory factor analysis was conducted in Mplus to identify underlying dimensions across the five DNAm clock variables (Horvath, Horvath Skin and Blood, Hannum, PhenoAge, and GrimAge). The number of factors to be rotated was determined by the scree plot of Eigenvalues (Supplementary Figure 2) and a likelihood ratio test comparing a two-factor model to a one-factor model. EFA models were estimated using full-information maximum likelihood with an oblique (geomin) rotation. Factor loadings were evaluated for statistical significance, with a criterion of p < .05 (Supplementary Table 3). We estimated composite scores based on the results of the EFA by standardizing the DNAmAge variables and summing the standardized values for the variables that loaded on each factor. As described in the results, the composite scores for each factor are referred to as Gen 1 and Gen 2. Each of the following models was fit for Gen 1 then repeated for Gen 2.

We first estimated univariate ACE models for Gen 1 and Gen 2 DNAmAge, decomposing their variances into genetic, shared environmental, and nonshared environmental components. Next, to explore the primary aims of the study, the relationship between DNAmAge and ΔIQ was modeled in two ways. Both the phenotypic association model and “quasi-causal” regression model were fit in Mplus. The phenotypic association model is equivalent to a simple phenotypic regression where the relationship between predictor and outcome is examined without controlling for genetic and shared environmental effects. This was estimated prior to all genetically informed models. Using a dataset wide by twin pair and estimated parameters set to be equal across the members of a pair, we regressed adult IQ on childhood IQ, childhood SES, DNAmAge, and the interaction between SES and DNAmAge.

Finally, we fit the quasi-causal regression model to estimate the phenotypic regression coefficients conditional on b_A and b_C paths. The classical twin method was used to partition childhood IQ, adult IQ, and DNAmAge into A, C, and E components. In this case, adult IQ was regressed on childhood IQ, DNAmAge, and the A and C components of DNAmAge and childhood IQ, as illustrated in Figure 1. Childhood SES was used as a linear covariate for DNAmAge and the IQ variables. We also estimated an interaction between childhood SES and DNAmAge to determine whether the strength of the relationship between DNAmAge and ΔIQ depended on the SES of the family in which twins were raised. A statistically significant b_P3 coefficient affirmatively answers the question: Within a pair of identical twins, does the twin with the higher DNAmAge value also have the larger positive or negative ΔIQ, statistically adjusting for the effects of the A and C components of DNAmAge (b_A3 and b_C3).