DunedinPACE, a DNA methylation biomarker of the pace of aging

  1. Daniel W Belsky  Is a corresponding author
  2. Avshalom Caspi
  3. David L Corcoran
  4. Karen Sugden
  5. Richie Poulton
  6. Louise Arseneault
  7. Andrea Baccarelli
  8. Kartik Chamarti
  9. Xu Gao
  10. Eilis Hannon
  11. Hona Lee Harrington
  12. Renate Houts
  13. Meeraj Kothari
  14. Dayoon Kwon
  15. Jonathan Mill
  16. Joel Schwartz
  17. Pantel Vokonas
  18. Cuicui Wang
  19. Benjamin S Williams
  20. Terrie E Moffitt
  1. Department of Epidemiology & Butler Columbia Aging Center, Columbia University, United States
  2. Center for Genomic and Computational Biology, Duke University, United States
  3. Department of Psychology and Neuroscience, Duke University, United States
  4. Department of Psychology, University of Otago, New Zealand
  5. Social, Genetic, and Developmental Psychiatry Centre, King's College London, United Kingdom
  6. Department of Environmental Health Sciences, Columbia University, United States
  7. Department of Occupational and Environmental Health, Peking University, China
  8. Complex Disease Epigenetics Group, University of Exeter, United Kingdom
  9. Robert N Butler Columbia Aging Center, Columbia University, United States
  10. Department of Environmental Health Sciences, Harvard TH Chan School of Public Health, Harvard University, United States
  11. Department of Medicine, VA Boston Healthcare System, United States

Abstract

Background:

Measures to quantify changes in the pace of biological aging in response to intervention are needed to evaluate geroprotective interventions for humans. Previously, we showed that quantification of the pace of biological aging from a DNA-methylation blood test was possible (Belsky et al., 2020). Here, we report a next-generation DNA-methylation biomarker of Pace of Aging, DunedinPACE (for Pace of Aging Calculated from the Epigenome).

Methods:

We used data from the Dunedin Study 1972–1973 birth cohort tracking within-individual decline in 19 indicators of organ-system integrity across four time points spanning two decades to model Pace of Aging. We distilled this two-decade Pace of Aging into a single-time-point DNA-methylation blood-test using elastic-net regression and a DNA-methylation dataset restricted to exclude probes with low test-retest reliability. We evaluated the resulting measure, named DunedinPACE, in five additional datasets.

Results:

DunedinPACE showed high test-retest reliability, was associated with morbidity, disability, and mortality, and indicated faster aging in young adults with childhood adversity. DunedinPACE effect-sizes were similar to GrimAge Clock effect-sizes. In analysis of incident morbidity, disability, and mortality, DunedinPACE and added incremental prediction beyond GrimAge.

Conclusions:

DunedinPACE is a novel blood biomarker of the pace of aging for gerontology and geroscience.

Funding:

This research was supported by US-National Institute on Aging grants AG032282, AG061378, AG066887, and UK Medical Research Council grant MR/P005918/1.

Editor's evaluation

This paper presents a new DNA methylation-based biomarker of aging: DunedinPACE. This biomarker is an updated version of DunedinPOAM, which was designed by the same group of authors to track an individual's Pace of Aging. It takes into account an additional measurement occasion (collected 20 years after inclusion) and only includes the most reliable DNA methylation probes, i.e. probes with little variation between technical replicates. DunedinPACE shows improved performance when compared to DunedinPOAM and can be used to complement previously generated DNA methylation-based biomarkers, such as GrimAge.

https://doi.org/10.7554/eLife.73420.sa0

Introduction

Guided by research on the cellular hallmarks of aging, model-organism studies are yielding treatments that slow the accumulation of deterioration in cells and organs (Barzilai et al., 2018; Bellantuono, 2018; Campisi et al., 2019). However, human geroprotector trials face a barrier when it comes to outcome measures: humans’ long lifespan makes it prohibitively time-consuming to test whether a treatment extends healthspan (Justice et al., 2018). What is needed is a way to measure each clinical trial participant’s personal pace of biological aging, before, during, after, and at long-term follow-up, to test whether a gero-protective therapy slows that pace, and whether benefits fade out (Belsky et al., 2017; Moffitt et al., 2017; Sierra et al., 2021).

An ideal research design for isolating biological patterns that differentiate faster and slower aging is longitudinal follow-up of a cohort of individuals who are all the same chronological age. We originally applied this design to test the geroscience-derived hypothesis that the process of biological aging was ongoing in healthy young adults and, critically, that it was already variable across individuals decades before most chronic diseases onset (Belsky et al., 2015; Elliott et al., 2021b). We studied a single-year birth cohort of all 1037 babies born in the city of Dunedin, New Zealand, during 1972–1973 and followed-up through midlife. Because aging can be understood as gradual, progressive deterioration simultaneously affecting different organ systems, we tracked declines in the cardiovascular, metabolic, renal, hepatic, immune, periodontal, and pulmonary systems of each participant using multiple biomarker measurements of each organ system, first at ages 26, 32, and 38 years, and subsequently including a fourth assessment at age 45 years. Modeling this one-of-a-kind dataset yielded a metric which quantified how slowly or rapidly each participant had been aging. We called this metric the Pace of Aging.

People with faster Pace of Aging tend to experience more rapid aging-related decline in physical and cognitive functions and develop early signs of brain aging (Belsky et al., 2015; Elliott et al., 2021a; Elliott et al., 2021b; Rasmussen et al., 2019). This evidence suggests that intervention to slow Pace of Aging could preserve functions lost with aging and extend healthspan. However, Pace of Aging takes many years to measure and requires data from several types of clinical and biological assays, limiting its utility in clinical trials. To address this limitation, we applied machine-learning tools to genome-wide DNA methylation data from Dunedin cohort members’ blood samples to distill the multi-timepoint, multi-assay Pace of Aging measure into a single-timepoint, single-assay blood biomarker previously described in this journal: DunedinPoAm (for Dunedin Study Pace of Aging from methylation; Belsky et al., 2020). DunedinPoAm was designed to measure Pace of Aging biological change over time from a single blood sample. Like the original Pace of Aging, people with faster DunedinPoAm scores more often experienced declines in cognitive and physical functioning by midlife and showed more rapid facial aging Belsky et al., 2020; in older adults, faster DunedinPoAm predicted increased risk of disease and death (Belsky et al., 2020; Graf et al., 2021); in young people, experiences of early-life adversity were linked to faster DunedinPoAm (Belsky et al., 2020; Raffington et al., 2021).

DunedinPoAm provided proof of concept that Pace of Aging could be measured from a single blood test. But DunedinPoAm was based on biological change observed over just twelve years spanning the transition from young adulthood to midlife, limiting the scope of biological change that could be observed. And this biological change was measured from just three assessment waves, limiting precision. In addition, similar to other DNA methylation composites measured from microarray data, the measure had only moderate test-retest reliability (Higgins-Chen et al., 2021), limiting its value to clinical trials seeking to test within-participant changes from pre-treatment baseline to post-treatment follow-up. To address these limitations, we incorporated new data from the Dunedin cohort extending Pace of Aging follow-up to include a fourth measurement occasion in the fifth decade of life (Elliott et al., 2021b) and restricted DNA-methylation data to exclude probes with poor reliability (Sugden et al., 2020) to refine our DNA-methylation biomarker of Pace of Aging. The longer follow-up period, which allowed us to observe more aging-related change in system integrity, and the additional measurement occasion included in analysis improve the power and precision of our modeling of longitudinal change in system integrity (Brandmaier et al., 2018; Rast and Hofer, 2014). The refined DNA-methylation dataset used for machine-learning improves reliability of measurement. We name the novel algorithm DunedinPACE, for Dunedin (P)ace of (A)ging (C)alculated from the (E)pigenome.

As a new DNA methylation measure of biological aging, DunedinPACE joins a well-established battery of DNA methylation measures known as ‘clocks’, so-named for their accurate prediction of chronological age (Horvath and Raj, 2018). However, DunedinPACE is distinct from the DNA methylation clocks in the method of its design and in its interpretation. DunedinPACE is derived from analysis of longitudinal data collected from a cohort of individuals who are all the same chronological age. It reflects differences between those individuals in the rate of deterioration in system integrity occurring over a fixed time interval, age 26 to age 45 years. It therefore has three distinguishing features that isolate aging from other confounds: (1) analysis of a single-year birth cohort distinguishes aging from confounding effects of exposures to factors that alter the methylome and that differ across generations, referred to hereafter as ‘cohort effects’; (2) initiation of follow-up in young adults distinguishes aging from the effects of disease and excludes contamination from survival bias; and (3) focus on changes in multi-organ-system integrity during adult life distinguishes ongoing aging processes from deficits established early in development. The clocks, in contrast, were developed using a single cross-section of data from samples of mixed-age and older adults. They do not distinguish aging from cohort effects, are vulnerable to survival bias and contamination of aging signal by disease processes, and do not distinguish aging-related changes from prenatal/childhood deficits. DunedinPACE, therefore, represents a complementary tool with useful advantages for DNA methylation quantification of biological aging.

Results

We developed the DunedinPACE measure reported here by analyzing the pace of biological aging in a 1972–1973 birth cohort (N = 1037), the Dunedin Study. This analysis consisted of two parts. The first part of analysis followed our original method (Belsky et al., 2015) to quantify Pace of Aging from two decades of longitudinal organ-system integrity data (Elliott et al., 2021b). First, we measured longitudinal changes in 19 biomarkers assessing cardiovascular, metabolic, renal, hepatic, immune, dental, and pulmonary systems, measured at ages 26, 32, 38 and, most recently in 2019, 45 years (Supplementary file 1A provides measurement details about the tracked biomarkers). Second, linear mixed-effects modeling was used to fit growth models to estimate each study member’s personal rate of change for each of the 19 biomarkers. Third, these 19 personal rates of change were combined across biomarkers to calculate each Study member’s personal Pace of Aging. In line with the geroscience hypothesis (Kaeberlein, 2013; Kennedy et al., 2014), which states that aging represents progressive decline across organ systems, we calculated each Study member’s Pace of Aging as the sum of age-dependent annual changes across all biomarkers: Pace of Agingi = B=119 μ1iB. The resulting Pace of Aging was then scaled to a mean of 1, so that it could be interpreted with reference to an average rate of 1 year of biological aging per year of chronological aging. Study members showed wide variation in their Pace of Aging (Mean = 1 biological year per chronological year, SD = 0.29). Over the two decades that we measured biological aging, the Study member with the slowest Pace of Aging aged by just 0.40 biological years per chronological year, while the Study member with the fastest Pace of Aging accrued 2.44 biological years per chronological year (Elliott et al., 2021b).

The second part of the analysis followed our original method to distill Pace of Aging into a blood test (Belsky et al., 2020). Using Illumina EPIC array DNA methylation data from blood collected at age 45, we developed DunedinPACE to provide a surrogate for the 20-year Pace of Aging. Briefly, we used elastic-net-regression (Zou and Hastie, 2005) to develop a DNA methylation algorithm to predict the 20-year Pace of Aging. Analysis included the subset of probes included on both the Illumina 450 k and EPIC arrays which we previously determined to have acceptable test-retest reliability (ICC >0.4, n = 81,2392) (Sugden et al., 2020). Following the method established by Horvath, 2013 and used in our prior analysis, we fixed the alpha parameter of the elastic net at 0.5. The final lamda value selected by the analysis was 0.042. The resulting algorithm included 173 CpG sites. The CpGs included in DunedinPACE are listed in Supplementary file 1B.

DunedinPACE distills the 20-year pace of aging into a single-timepoint DNA methylation blood test

DunedinPACE demonstrated a high in-sample correlation with Pace of Aging to age 45 (r = 0.78, Figure 1a). The value indicates improved fit of the DunedinPACE model to the 20-year Pace of Aging as compared to the original DunedinPoAm version’s fit to the 12-year Pace of Aging (in-sample correlation r = 0.6). Correlations among the 12- and 20-year Pace of Aging measures and the DunedinPoAm and DunedinPACE DNA methylation measures are shown in Figure 1—figure supplement 1.

Figure 1 with 4 supplements see all
DunedinPACE correlation with 20-year Pace of Aging and association with decline in physical and cognitive functions and subjective signs of aging.

Panel A shows the association of DunedinPACE with the 20-year Pace of Aging in the Dunedin Study cohort (r = 0.78). Panel B shows effect-sizes for associations of DunedinPACE with measures of physical and cognitive functioning and subjective signs of aging in Dunedin Study members at age 45 years. Colors indicate groupings of measures (physical functions in shades of orange, cognitive functions in shades of yellow, subjective signs of aging in shades of green). Stars next to labels for Physical Limitations and Facial Aging indicate that these meaures were reverse-coded for this analysis. Panel C shows binned scatterplots of associations of DunedinPACE with declines in physical and cognitive function and subjective signs of aging. Plotted points show average values for ‘bins’ of approximately 20 Study members. Regression slopes are estimated from the raw, un-binned data. In Panel C, y-axis scales are denominated in units of standard deviations computed from the baseline measurement for all outcomes except self-rated health, for which the y-axis shows probability of incident fair/poor self-rated health. Changes were calculated over the interval between age-38 and age-45 assessments for all measures except cognition, for which change was calculated over the interval between the age-13 and age-45 assessments. DunedinPACE and DunedinPoAm were developed from analysis of Pace of Aging in the Dunedin cohort; effect-sizes for these measures may be over-estimated.

Consistent with the high correlation of DunedinPACE with 20-year Pace of Aging, effect-sizes for associations of DunedinPACE with measures of physical and cognitive functioning and subjective signs of aging measured at age 45 in the Dunedin Study were similar to effect-sizes for the 20-year Pace of Aging. Effect-sizes for DunedinPACE are graphed in Figure 1b. Comparison with effect-sizes for the 20-year Pace of Aging reported previously (Elliott et al., 2021b; Rasmussen et al., 2019) are shown in Figure 1—figure supplement 2.

We also repeated analysis of decline in physical and cognitive functions included in our original 2020 eLife report of DunedinPoAm. This analysis tested associations between DunedinPACE and worsening in physical functions and subjective signs of aging between the age-38 and age-45 assessments and change in cognitive functioning from adolescent baseline to the age-45 assessment. Effect-sizes are graphed in Figure 1c.

Comparison of DunedinPACE effect-sizes for associations with function tests and subjective signs of aging to effect-sizes for the original DunedinPoAm and the DNA methylation clocks proposed by Horvath, Hannum et al., Levine et al. (PhenoAge Clock), and Lu et al. (GrimAge Clock) are shown in Figure 1—figure supplements 3 and 4. DunedinPACE effect-sizes are similar to or larger than those for DunedinPoAm and the GrimAge clock and larger than those for the other DNA methylation clocks. DunedinPACE and DunedinPoAm were developed from analysis of Pace of Aging in the Dunedin cohort; effect-sizes for these measures may be over-estimated in analyses of Dunedin data. We therefore conducted further validation analysis in four additional cohorts.

DunedinPACE shows exceptional test-retest reliability

A Pace of Aging measure should have high test-retest reliability in order to be able to quantify change from pre-treatment baseline to post-treatment follow-up in intervention studies or exposure-related changes in longitudinal observational studies. We first evaluated test-retest reliability of DunedinPACE using a publicly available database of 36 replicate adult-blood-sample Illumina 450 k datasets (Lehne et al., 2015) (GEO Accession GSE55763). We estimated intraclass correlation coefficients (ICCs) for these replicates using mixed-effects regression. For the original DunedinPoAm, reliability was good (ICC = 0.89 95% CI [0.79–0.94]). For DunedinPACE, reliability was excellent (ICC = 0.96 [0.93–0.98]; Figure 2).We repeated this analysis in two additional databases of replicates reported by Sugden et al., 2020, a database of n = 28 replicate Illumina EPIC-array datasets and a database of n = 350 replicates in which one dataset was generated from the Illumina 450 k array and the other dataset was generated from the Illumina EPIC array. DunedinPACE ICCs were 0.97 [0.94–0.98] in the EPIC-EPIC database and 0.87 [0.82–0.90] in the 450k-EPIC database.

Figure 2 with 1 supplement see all
Test-retest reliability of DunedinPACE.

The figure graphs DunedinPACE values for replicate Illumnina450k datasets for n = 36 individuals in the dataset published by Lehne and colleagues (Lehne et al., 2015) (GEO Accession GSE55763). The ICC for DunedinPACE in the Lehne dataset is 0.96 95% CI [0.92–0.98].

Comparison of DunedinPACE reliability to reliabilities for the original DunedinPoAm and the DNA methylation clocks are shown in Figure 2—figure supplement 1. DunedinPACE is as reliable as the GrimAge clock and more reliable in comparison to the other clocks, although methods have been proposed that may improve reliabilities for the other clocks (Higgins-Chen et al., 2021).

We conducted analysis to evaluate the extent to which improved fit to longitudinal Pace of Aging and better test-retest reliability for DunedinPACE as compared to the original DunedinPoAm might reflect (a) the four-time-point follow-up design of 20-year Pace of Aging and (b) the restriction of elastic net regression to develop DunedinPACE to probes that demonstrated higher test-retest reliabilities in blood in an independent sample. Results suggest that both four-time-points to measure Pace of Aging and restricting machine-learning analysis to CpG sites with higher test-retest reliability contributed to the technical improvement in DunedinPACE over DunedinPoAm (see Appendix 1).

DunedinPACE is faster in chronologically and biologically older individuals

In demography, the pace of biological aging in a population can be estimated from the rate of increase in mortality risk from younger to older chronological ages (Hägg et al., 2019). In humans and many other species, the increase in mortality risk with advancing chronological age is nonlinear, suggesting that the pace of biological aging accelerates as we grow older (Gompertz, 1997; Kirkwood, 2015; Olshansky and Carnes, 1997). We tested if, consistent with this hypothesis, DunedinPACE was faster in chronologically older as compared to younger individuals in data from the Understanding Society Study (n = 1175,, age range 28–95). Chronologically older Understanding Society participants had faster DunedinPACE as compared to younger ones (r = 0.32, Figure 3A). This correlation was threefold larger as compared to the correlation for the original DunedinPoAm (r = 0.11).

Figure 3 with 1 supplement see all
Associations of DunedinPACE with chronological age, epigenetic clocks, Physiology-based measures of biological age, and self-rated health in Understanding Society.

Panel A shows a scatterplot and fitted slopes illustrating the association between chronological age (x-axis) and DunedinPACE (y-axis) in women and men in the Understanding Society sample. Data for women are plotted with yellow dots (orange slope) and for men with blue crosses (navy slope). The figure illustrates a positive association between chronological age and DunedinPACE (Pearson r = 0.32). Panel B shows a matrix of correlations and association plots among DunedinPACE and age-acceleration residuals of Horvath, Hannum, Levine-PhenoAge and Lu-GrimAge epigenetic clocks. The diagonal cells of the matrix list the DNA methylation measures. The half of the matrix below the diagonal shows scatter plots of associations. For each scatter-plot cell, the y-axis corresponds to the variable named along the matrix diagonal to the right of the plot and the x-axis corresponds to the variable named along the matrix diagonal above the plot. The half of the matrix above the diagonal lists Pearson correlations between the DNA methylation measures. For each correlation cell, the value reflects the correlation of the variables named along the matrix diagonal to the left of the cell and below the cell. Panel C graphs scatterplots of associations of DunedinPACE with three physiology-based measures of biological age (KDM Biological Age Advancement, r = 0.30 [0.24–0.36]; Phenotypic Age Advancement, r = 0.32 95% CI [0.26–0.38]; and Homeostatic Dysregulation r = 0.09 [0.03–0.16]) and a plot of DunedinPACE means and 95% confidence intervals by self-rated health category (r = 0.20 [0.15–0.26]).

Between-individual correlation of older chronological age with faster DunedinPACE could be confounded by differences in exposure histories across birth cohorts, that is not an effect of within-individual aging (Moffitt, 2020). To rule out such confounding, we conducted analysis of within-individual change in DunedinPACE in the US Veterans Administration Normative Aging Study (NAS). Among NAS participants with repeated-measures DNA methylation data (n = 536), DunedinPACE increased by b = 0.021 (SE = 0.003) per 5 years of follow-up, or approximately 0.2 standard deviation units, similar to what we previously reported for the original DunedinPoAm (Belsky et al., 2020).

DunedinPACE is faster in individuals measured to be biologically older using prior measures derived from DNA methylation and physiological data, and who report poorer self-rated health. We compared DunedinPACE with the epigenetic clocks proposed by Hannum et al., 2013; Horvath, 2013; Levine et al., 2018; Lu et al., 2019, with measurements of biological age derived from physiological data and with participants’ subjective perceptions of their own health status in data from the UK-based Understanding Society Study (n = 1175,, mean age = 58, SD = 15, 42% male).

Epigenetic clock analysis

For each epigenetic clock, we calculated the measure referred to as epigenetic ‘age acceleration’ by regressing participants’ clock-estimated ages on their chronological ages and computing residual values. This measure is often used as an estimate of the aging rate. It quantifies the difference between the amount of aging expected based on a person’s chronological age and the amount of aging observed based on DNA methylation. DunedinPACE was weakly correlated with age acceleration computed from the Horvath clock (r = 0.13) and somewhat more strongly correlated with age acceleration computed from the Hannum and Levine clocks (Hannum r = 0.26; PhenoAge r = 0.36). These correlations were similar to those reported on our 2020 eLife article for the original DunedinPoAm measure. For the GrimAge clock proposed by Lu et al., the correlation with the age-acceleration residuals was stronger (r = 0.58). Correlations are plotted in Figure 3B.

Physiology-based measurements of biological age

We used published methods to measure participants’ biological ages from blood-chemistry, blood-pressure, and lung function data collected at the same time as the DNA methylation data. We computed indices of biological age using three different methods: the Klemera and Doubal method (KDM) Biological Age estimates the age at which a person’s physiology would appear normal in a reference population Klemera and Doubal, 2006; the Phenotypic Age method estimates the age at which a person’s physiology-predicted mortality risk would be approximately normal in a reference population Levine et al., 2018; the homeostatic dysregulation method estimates the extent to which a person’s physiology deviates from a state of health (Cohen et al., 2013). These algorithms generate correlated but distinct estimates conceptually related to biological age (Hastings et al., 2019; Parker et al., 2020). Algorithms were parameterized using a set of biomarkers available in Understanding Society and included in Levine’s original implementation of the Klemera-Doubal Biological Age (Levine, 2013; albumin, alkaline phosphatase, blood urea nitrogen, creatinine, C-reactive protein, HbA1C, systolic blood pressure, forced expiratory volume in one second) and training data from the US National Health and Nutrition Examination Surveys (NHANES) (Kwon and Belsky, 2021). Consistent with analysis of epigenetic age acceleration, participants with more advanced biological aging measured from physiological data also showed faster DunedinPACE (KDM Biological Age Advancement r = 0.30; Phenotypic Age Advancement r = 0.32; Homeostatic Dysregulation r = 0.09; Figure 3C).

Self-rated health

Participants who rated themselves as being in worse health also showed faster DunedinPACE. The difference in DunedinPACE between those rating their health as excellent and poor was large (Cohen’s d = 0.74 [0.46–1.03]; Figure 3C).

Comparison of effect-sizes for associations with physiology-based measures of biological age and self-rated health between DunedinPACE and the original DunedinPoAm and the four DNA methylation clocks is shown in Figure 3—figure supplement 1. DunedinPACE effect-sizes are similar effect-sizes for DunedinPoAm and the GrimAge clock and larger in comparison to the other clocks.

Faster DunedinPACE is associated with morbidity, disability, and mortality

We conducted analysis of morbidity and mortality in older men in the VA Normative Aging Study (NAS) and in the Framingham Heart Study Offspring cohort.

NAS analysis was conducted using the same methods reported in our original article (Belsky et al., 2020). Briefly, analysis included n = 771 older men with average age of 77 years (SD = 7) at DNA methylation measurement. Over follow-up from DNA collection in 1999–2013, 46% died and 23% were diagnosed with a new chronic disease, including any of hypertension, type-2 diabetes, cardiovascular disease, chronic obstructive pulmonary disease, chronic kidney disease, and cancer. Participants with faster DunedinPACE were at increased risk for morbidity and mortality (for incident chronic disease morbidity, DunedinPACE HR = 1.23 [1.07–1.42]; for prevalent chronic disease morbidity, DunedinPACE RR = 1.16 [1.12–1.20]; for mortality, DunedinPACE HR = 1.26 95% CI [1.14–1.40], Figure 4A).

Figure 4 with 4 supplements see all
Association of DunedinPACE with mortality.

Panel A shows mortality in the Normative Aging Study (NAS). Panel B shows mortality in the Framingham Heart Study Offspring Cohort. The figure plots Kaplan-Meier curves for three groups of participants in each of the two cohorts: those with DunedinPACE 1 SD or more below the mean (‘slow’ DunedinPACE, blue line); those with DunedinPACE within 1 SD of the mean (‘average’ DunedinPACE, purple line); and those with DunedinPACE 1 SD or more above the mean (‘fast’ DunedinPACE, red line). Censoring of participants prior to death is indicated with a gold hash marks. The table below the figure details the number of participants at risk per 3-year interval and, in parentheses, the number who died during the interval.

Framingham analysis included n = 2471 members of the Offspring cohort (54% women) with average age of 66 years (SD = 9) at DNA methylation measurement during 2005–2008. Over follow-up through 2018, 23% died, 13% were newly diagnosed with cardiovascular disease (CVD), and 6% had a first stroke or transient ischemic attack (TIA). Disability follow-up was conducted through 2015 based on participant reports of limitations to activities of daily living (ADLs) on the Nagi, Katz, and Rosow-Brelsau scales. Participants with faster DunedinPACE at baseline were at increased risk for CVD, stroke/TIA, and mortality (CVD HR = 1.39 [1.26–1.54]; stroke/TIA HR = 1.37 [1.19–1.58]; mortality HR = 1.65 95% CI [1.51–1.79], Figure 4B). They were also more likely to develop disability (Nagi ADL IRR = 1.40 [1.19–1.65]; Katz ADL IRR = 1.33 [1.16–1.53]; Rosow-Breslau ADL IRR = 1.39 [1.24–1.56]).

Comparison of effect-sizes for associations with morbidity, disability, and mortality between DunedinPACE and the original DunedinPoAm and the DNA methylation clocks are shown in Figure 4—figure supplements 14. DunedinPACE effect-sizes were similar to those for DunedinPoAm and the GrimAge clock and larger than effect-sizes for the other measures.

Childhood exposure to poverty and victimization is associated with faster DunedinPACE

A Pace of Aging measure should be sensitive to histories of exposure associated with earlier onset of morbidity and shorter lifespan and should provide indications of faster aging before clinical signs of aging-related disease develop, in time for prevention. To test if DunedinPACE indicated faster aging in young people with histories of exposure thought to shorten healthy lifespan, we analyzed data from n = 1,658 members of the E-Risk Longitudinal Study. The E-Risk Study follows a 1994–1995 birth cohort of same-sex twins. Blood DNA methylation data were collected when participants were aged 18 years. We analyzed two exposures associated with shorter healthy lifespan, childhood low socioeconomic status and childhood victimization. Socioeconomic status was measured from data on parents’ education, occupation, and income (Trzesniewski et al., 2006). Victimization was measured from exposure dossiers compiled from interviews with the children’s mothers and home-visit assessments conducted when the children were aged 5, 7, 10, and 12 (Fisher et al., 2015). The dossiers recorded children’s exposure to domestic violence, peer bullying, physical and sexual harm by an adult, and neglect. 72% of the analysis sample had no victimization exposure, 21% had one type of victimization exposure, 4% had two types of exposure, and 2% had three or more types of exposure.

E-Risk participants with exposure to childhood poverty and victimization showed faster DunedinPACE (For low as compared to high childhood socioeconomic status, d = 0.38 95% CI [0.25–0.51]; for polyvictimization as compared to no victimization, d = 0.47 [0.17–0.77]; Figure 5).

Figure 5 with 1 supplement see all
DunedinPACE levels by strata of childhood socioeconomic status (SES) and victimization in the E-Risk Study.

Panel A (left side) plots means and 95% CIs for DunedinPACE measured at age 18 among E-Risk participants who grew up low, middle, and high socioeconomic status households. Panel B (right side) plots means and 95% CIs for DunedinPACE measured at age 18 among E-Risk participants who experienced 0, 1, 2, or 3 or more types of victimization through age 12 years.

Comparison of effect-sizes for associations with childhood poverty and victimization between DunedinPACE and the original DunedinPoAm and the DNA methylation clocks are shown in Figure 5—figure supplement 1. DunedinPACE effect-sizes were similar to those for DunedinPoAm and the GrimAge clock and larger than effect-sizes for the other clocks.

Sensitivity analysis

Results for analysis of all studies including covariate adjustment for cell counts and smoking are reported in Supplementary file 2. Covariate adjustment for estimated cell counts and participant reports of smoking history resulted in modest attenuation of some effect-sizes, although DunedinPACE effect-sizes remained statistically different from zero in nearly all cases (associations with declines in balance and cognitive function were attenuated below the alpha = 0.05 threshold in Dunedin Study analysis including smoking as a covariate; differences between un-victimized and those with one type of victimization were attenuated below the alpha = 0.05 threshold in E-Risk Study analysis including smoking as a covariate; differences between high- and middle-SES groups were attenuated below the alpha = 0.05 threshold in E-Risk Study analysis excluding smokers). Compared to the original DunedinPoAm, covariate adjustment for smoking tended to have a smaller impact on effect-size estimates for DunedinPACE. In follow-up analysis of the Understanding Society Study, we found that participant-reported smoking history explained 30% of variance in DunedinPoAm, but only 13% of variance in DunedinPACE. Results were similar in analysis of the Framingham Heart Study Offspring Cohort (R2 = 23% for DunedinPoAm vs. 12% for DunedinPACE).

Incremental predictive value of DunedinPACE in an analysis of morbidity, disability, and mortality

DunedinPACE is distinct from DNA methylation clocks in the method of its development and in its interpretation. However, DunedinPACE is moderately correlated with some DNA methylation clocks, in particular GrimAge (see Figure 3B). We therefore conducted analysis to test if DunedinPACE contributed new information about health-span and lifespan over and above existing DNA methylation clocks. We tested if DunedinPACE associations with morbidity, disability, and mortality were statistically independent of each of the DNA methylation clocks within the Framingham Heart Study. In the case of the GrimAge clock, which was developed to predict mortality using this Framingham dataset, this analysis provides an especially rigorous test.

DunedinPACE associations with health-span endpoints were little-changed by covariate adjustment for the Horvath, Hannum, or PhenoAge Clocks. In models adjusted for GrimAge, DunedinPACE associations with mortality, CVD, and disability were attenuated, but remained statistically different from zero (mortality HR = 1.24 [1.49–1.74], CVD HR = 1.18 [1.05–1.34], Nagi ADL IRR = 1.27 [1.02–1.58], Katz ADL IRR = 1.26 [1.02–1.54], Rosow-Breslau ADL IRR = 1.27 [1.08–1.50]); associations with stroke were similar to unadjusted models (HR = 1.33 [1.05–1.69]). Results for all models are reported in Supplementary file 1C. Thus, Dunedin PACE adds incremental prediction over and above all clocks studied here.

Discussion

We developed a novel DNA methylation measure of the pace of biological aging from analysis of two decades of longitudinal data from members of the Dunedin Study 1972–1973 birth cohort. The new measure is named DunedinPACE. As in our original report (Belsky et al., 2020), we utilized a two-step approach that first modeled the Pace of Aging from within-individual changes in organ-system integrity markers and then distilled this longitudinal change measure into a single-time-point DNA-methylation blood test. DunedinPACE is therefore distinct from the DNA methylation clocks in both theory and method. The clocks estimate the progress of aging up to a cross-section in time. DunedinPACE is a DNA methylation estimate of the Pace of Aging, the ongoing rate of decline in system integrity.

The DunedinPACE measure advances DNA methylation measurement of the Pace of Aging for clinical trials of geroprotective therapies beyond our original DunedinPoAm measure in four ways: (1) Pace of Aging analysis included 20 years of follow-up (instead of 12); (2) Pace of Aging analysis included four time points of measurement (instead of three); (3) DNA methylation modeling of Pace of Aging excluded CpG sites for which probes were determined to have low test-retest reliability in blood datasets; (4) the algorithm to implement DunedinPACE includes a normalization step that allows DunedinPACE values for individual samples to be compared to the Dunedin Study cohort in which the measure was developed.

As a result of these advances, DunedinPACE provides a more precise measurement of the Pace of Aging as compared to the original DunedinPoAm and shows higher test-retest reliability. Interestingly, in contrast to reports for some epigenetic clocks (Higgins-Chen et al., 2021), the restriction of machine learning analysis to a more-reliably-measured subset of CpG sites on the Illumina EPIC array did not harm criterion validity of DunedinPACE; effect-sizes for prediction of morbidity and mortality were as large or larger in comparison to DunedinPoAm. Analysis of this more precise and reliable measurement revealed stronger associations with signs of aging. DunedinPACE showed a stronger correlation with chronological age as compared to the original DunedinPoAm, consistent with demographic evidence that biological processes of aging accelerate in later life (Finch and Crimmins, 2016). In analysis of construct and criterion validity, DunedinPACE was associated with functional decline in midlife adults; it was correlated with measures of biological age derived from blood chemistry and DNA methylation data, and with research participants’ subjective perceptions of their own health; it was predictive of morbidity and mortality in midlife and older adults; and it indicated faster Pace of Aging in young adults with histories of exposure to poverty and victimization. DunedinPACE effect-sizes were similar to or larger than those for the original DunedinPoAm measure even as DunedinPACE was less influenced by participant smoking history. In analysis that took smoking history into account, DunedinPACE effect-sizes were more robust than those for DunedinPoAm, most notably in analysis of mortality in the Framingham Heart Study and of early-life adversity in the E-Risk cohort.

DunedinPACE effect-sizes in analysis of construct and criterion validity were also larger than those for benchmark DNA methylation clocks proposed by Horvath, Hannum et al., and Levine et al. (PhenoAge), and similar to those for the clock proposed by Lu et al. (GrimAge), and were robust to covariate adjustment for smoking and DNA methylation estimates of cell counts. Moreover, in Framingham Heart Study Offspring Cohort analysis of incremental predictive validity, DunedinPACE associations with morbidity, disability, and mortality were robust to covariate adjustment for the DNA methylation clocks, including GrimAge, although some associations were attenuated in GrimAge-adjusted models. Framingham analysis represents a decisive test of value added by DunedinPACE over and above GrimAge because GrimAge was developed within the Framingham cohort.

In sum, DunedinPACE represents a novel measure of aging that is conceptually and empirically distinct from the DNA methylation clocks and that constitutes an advance on the original DunedinPoAm measure. DunedinPACE can complement existing DNA methylation measures of aging in analysis of observational studies and randomized trials to help advance the frontiers of geroscience.

We acknowledge limitations. Foremost, the Dunedin Study sample we analyzed to develop DunedinPACE is a relatively modestly sized cohort and is drawn from a single country. As additional cohorts develop the multi-time-point organ-system-function datasets to model Pace of Aging, integrating these data into DNA methylation algorithms to measure Pace of Aging may enhance precision. Nevertheless, DunedinPACE does show consistent evidence of criterion and construct validity in several additional datasets, boosting confidence in its external validity. In addition, our work on DunedinPACE thus far has not addressed population diversity in biological aging. The Dunedin cohort and the Understanding Society, NAS, Framingham, and E-Risk cohorts were mostly of white European descent. Follow-up of DunedinPACE in more diverse samples is needed to establish cross-population validity. Importantly, our original DunedinPoAm measure has been followed-up in diverse cohorts and has showed evidence of consistent validity across race/ethnic subgroups (Crimmins et al., 2021; Graf et al., 2021; Raffington et al., 2021; Schmitz et al., 2021). Criterion validity analyses conducted in this article do not consider specific diseases or causes of death because the cohorts used for follow-up are relatively small. Geroscience theory predicts aging should increase risk of chronic disease and death overall. Analysis of DunedinPACE in larger datasets with more deaths will allow for tests of whether particular incident diseases or causes of death drive associations. Comparative analyses in this article include only other DNA methylation measures of aging. Recent reports suggest valuable information about aging and healthspan may also be captured in metabolomic and proteomic datasets (Deelen et al., 2019; Eiriksdottir et al., 2021; Jansen et al., 2021; Lehallier et al., 2019; Tanaka et al., 2020). Future studies should compare DunedinPACE to measures derived from these biological levels of analysis. As with all proposed measures of biological aging, ultimately establishing DunedinPACE as a surrogate endpoint for healthspan will require evidence that it is modifiable by intervention and that intervention-induced changes result in changes in healthy-lifespan phenotypes (Justice et al., 2018; Prentice, 1989).

Within the bounds of these limitations, our analysis establishes DunedinPACE as a novel single-time-point measure that quantifies Pace of Aging from a blood test. It can be implemented in Illumina 450 k or in EPIC array data, making it immediately available for testing in a wide range of existing datasets as a complement to existing methylation measures of aging. Code to compute DunedinPACE is publicly available on GitHub (Corcoran et al., 2022). Through collaboration with ongoing cohort studies, DunedinPACE will soon be available from the US Health and Retirement Study and other studies. Because of its high test-retest reliability, DunedinPACE offers a unique measurement for intervention trials and natural experiment studies investigating how the rate of aging may be changed by behavioral or drug therapy, or by environmental modification.

Materials and methods

Data sources

Data were used from five studies: The Dunedin Study, the Understanding Society Study, the Normative Aging Study (NAS), the Framingham Heart Study, and the Environmental Risk (E-Risk) Longitudinal Twin Study. In addition, we accessed the Gene Expression Omnibus dataset GSE55763. The datasets and measures analyzed within each of them are described below.

The Dunedin Study

Request a detailed protocol

The Dunedin Study is a longitudinal investigation of health and behavior in a complete birth cohort. Study members (N = 1037; 91% of eligible births; 52% male) were all individuals born between April 1972 and March 1973 in Dunedin, New Zealand (NZ), who were eligible based on residence in the province and who participated in the first assessment at age 3. The cohort represents the full range of socioeconomic status on NZ’s South Island and matches the NZ National Health and Nutrition Survey on key health indicators (e.g. BMI, smoking, physical activity, GP visits) (Poulton et al., 2015). The cohort is primarily white (93%) (Poulton et al., 2015). Assessments were carried out at birth and ages 3, 5, 7, 9, 11, 13, 15, 18, 21, 26, 32, 38 and, most recently, 45 years, when 94% of the 997 study members still alive took part. At each assessment, each study member is brought to the research unit for a full day of interviews and examinations. Study data may be accessed through agreement with the Study investigators (https://moffittcaspi.trinity.duke.edu/research-topics/dunedin). Dunedin Study measures of physical and cognitive functioning and subjective signs of aging are described in detail in Supplementary file 1D.

GSE55763

Request a detailed protocol

GSE55763 is a publicly available dataset including technical replicate Illumina 450 k DNA methylation array data for 36 adult human samples (Lehne et al., 2015). Data are available from the Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE55763).

Understanding Society

Request a detailed protocol

Understanding Society is an ongoing panel study of the United Kingdom population (https://www.understandingsociety.ac.uk/). During 2010–2012, participants were invited to take part in a nurse’s exam involving a blood draw. Of the roughly 20,000 participants who provided clinical data in this exam, methylation data have been generated for just under 1200. We analyzed data from 1175 participants with available methylation and blood chemistry data. Documentation of the methylation (Yanchun, 2017) and blood chemistry (University of Essex, 2013) data resource is available online (here).

Physiology-based biological age measures
Request a detailed protocol

We measured biological aging from blood chemistry, systolic blood pressure, and lung-function data using the algorithms proposed by Klemera and Doubal, Levine, and Cohen (Cohen et al., 2013; Klemera and Doubal, 2006; Levine, 2013; Levine et al., 2018), trained in data from US National Health and Nutrition Examination Surveys (NHANES; https://wwwn.cdc.gov/nchs/nhanes/Default.aspx) following the methods described by Hastings (Hastings et al., 2019) and using the software developed by Kwon (https://github.com/dayoonkwon/BioAge, copy archived at swh:1:rev:bac0d617f4d04eb76350c0750cc367d8d1fd719e, Kwon, 2022). Following the procedure in our original eLife article (Belsky et al., 2020), we included 8 of 10 biomarkers included in Levine’s original implementation of the Klemera-Doubal method Biological Age (Levine, 2013): albumin, alkaline phosphatase (log), blood urea nitrogen, creatinine (log), C-reactive protein (log), HbA1C, systolic blood pressure, and forced expiratory volume in 1 second (FEV1). We omitted total cholesterol because of evidence this biomarker shows different directions of association with aging in younger and older adults (Arbeev et al., 2016). Cytomegalovirus optical density was not available in the Understanding Society database.

Self-rated health
Request a detailed protocol

Understanding Society participants rated their health as excellent, very-good, good, fair, or poor. We standardized this measure to have Mean = 0, Standard Deviation = 1 for analysis.

The Normative Aging Study

The Normative Aging Study (NAS) is an ongoing longitudinal study on aging established by the US Department of Veterans Affairs in 1963. Details of the study have been published previously (Bell et al., 1972). Briefly, the NAS is a closed cohort of 2280 male veterans from the Greater Boston area enrolled after an initial health screening to determine that they were free of known chronic medical conditions. Participants have been re-evaluated every 3–5 years on a continuous rolling basis using detailed on-site physical examinations and questionnaires. DNA from blood samples was collected from 771 participants beginning in 1999. We analyzed blood DNA methylation data from up to four repeated assessments conducted through 2013 (Gao et al., 2019; Panni et al., 2016). Of the 771 participants with DNA methylation data, n = 536 (70%) had data from two repeated assessments and n = 178 (23%) had data from three or four repeated assessments. We restricted the current analysis to participants with at least one DNA methylation data point. The NAS was approved by the Department of Veterans Affairs Boston Healthcare System and written informed consent was obtained from each subject before participation.

Mortality

Request a detailed protocol

Regular mailings to study participants have been used to acquire vital-status information and official death certificates were obtained from the appropriate state health department to be reviewed by a physician. Participant deaths are routinely updated by the research team and the last available update was on 31 December 2013. During follow-up, n = 354 (46%) of the 771 NAS participants died.

Chronic disease morbidity

Request a detailed protocol

We measured chronic disease morbidity from participants medical histories and prior diagnoses (Gao et al., 2021; Lepeule et al., 2018). We counted the number of chronic diseases to compose an ordinal index with categories of 0, 1, 2, 3, or 4+ of the following comorbidities: hypertension, type-2 diabetes, cardiovascular disease, chronic obstructive pulmonary disease, chronic kidney disease, and cancer.

The Framingham Heart Study

The Framingham Heart Study tracks the development of cardiovascular disease in three generations of families recruited in Framingham Massachusetts beginning in 1948 (Tsao and Vasan, 2015). We analyzed data from the Offspring Cohort, the second generation of study participants, who were recruited beginning in 1971. Blood samples used for DNA methylation analysis were collected from Offspring Cohort members at their 8th follow-up visit during 2005–2008. We analyzed data for n = 2471 cohort members with available DNA methylation and mortality data. This sample was 54% women and had an average age of 66 years (SD = 9) at the time blood draws for DNA methylation analysis were conducted. Data for the Framingham Study were obtained from dbGaP (phs000007.v32.p13).

Mortality

Request a detailed protocol

Mortality ascertainment in our dataset extended through 2018 (dbGaP accession pht003317.v9.p13, dataset vr_survdth_2018_a_1268 s). Participants contributed a maximum of 14 years of follow-up time. A total of 575 deaths were recorded at an average age of 81 years (SD = 9).

Cardiovascular disease (CVD)

Request a detailed protocol

CVD ascertainment in our dataset extended through 2018 (dbGaP accession pht003316.v9.p13, dataset vr_survcvd_2018_a_1267 s). N = 2,075 participants free of cardiovascular disease at baseline contributed a maximum of 14 years of follow-up time during which 327 incident cases of CVD were recorded at an average age of 77 years (SD = 6).

Stroke/transient ischemic attack (TIA)

Request a detailed protocol

Stroke/TIA ascertainment in our dataset extended through 2018 (dbGaP accession pht006024.v4.p13, dataset vr_svstktia_2018_a_1270 s). N = 2364 participants with no history of stroke/TIA at baseline contributed a maximum of 14 years of follow-up time during which 149 events were recorded at an average age of 78 years (SD = 8).

Disability

Request a detailed protocol

Disability was measured from participant reports about limitations to activities of daily living (ADLs) during interviews conducted at the 8th and 9th assessment waves. ADL information was collected using the Nagi, Katz, and Rosow-Breslau scales (Katz et al., 1970; Nagi, 1976; Rosow and Breslau, 1966). Scales are described in Supplementary file 1E. Incident disability was analyzed for participants free of disability at baseline (Nagi n = 1830; Katz n = 1826; Rosow-Breslau n = 1670). Over follow-up, n = 123 participants developed new limitations in Nagi ADLs; n = 181 developed new limitations in Katz ADLs; and n = 224 developed new limitations in Rosow-Breslau ADLs.

The Environmental Risk Longitudinal Study

The Environmental Risk Longitudinal Twin Study (E-Risk) tracks the development of a birth cohort of 2232 British participants. The sample was drawn from a larger birth register of twins born in England and Wales in 1994–1995. Full details about the sample are reported elsewhere (Moffitt and Team, 2002). Briefly, the E-Risk sample was constructed in 1999–2000, when 1116 families (93% of those eligible) with same-sex 5-year-old twins participated in home-visit assessments. This sample comprised 56% monozygotic (MZ) and 44% dizygotic (DZ) twin pairs; sex was evenly distributed within zygosity (49% male). Families were recruited to represent the UK population of families with newborns in the 1990s, on the basis of residential location throughout England and Wales and mother’s age. Teenaged mothers with twins were over-selected to replace high-risk families who were selectively lost to the register through non-response. Older mothers having twins via assisted reproduction were under-selected to avoid an excess of well-educated older mothers. The study sample represents the full range of socioeconomic conditions in the UK, as reflected in the families’ distribution on a neighborhood-level socioeconomic index (ACORN [A Classification of Residential Neighborhoods], developed by CACI Inc for commercial use): 25.6% of E-Risk families lived in ‘wealthy achiever’ neighborhoods compared to 25.3% nationwide; 5.3% vs 11.6% lived in ‘urban prosperity’ neighborhoods; 29.6% vs 26.9% lived in ‘comfortably off’ neighborhoods; 13.4% vs 13.9% lived in ‘moderate means’ neighborhoods, and 26.1% vs 20.7% lived in ‘hard-pressed’ neighborhoods. E-Risk underrepresents ‘urban prosperity’ neighborhoods because such households are likely to be childless.

Home-visits assessments took place when participants were aged 5, 7, 10, 12, and 18 years, when 93% of the participants took part. At ages 5, 7, 10, and 12 years, assessments were carried out with participants as well as their mothers (or primary caretakers); the home visit at age 18 included interviews only with participants. Each twin was assessed by a different interviewer. These data are supplemented by searches of official records and by questionnaires that are mailed, as developmentally appropriate, to teachers, and co-informants nominated by participants themselves. The Joint South London and Maudsley and the Institute of Psychiatry Research Ethics Committee approved each phase of the study. Parents gave informed consent and twins gave assent between 5 and 12 years and then informed consent at age 18. Study data may be accessed through the Study investigators (https://moffittcaspi.trinity.duke.edu/research-topics/erisk).

Childhood Socioeconomic Status

Request a detailed protocol

Childhood Socioeconomic Status (SES) was defined through a standardized composite of parental income, education, and occupation (Trzesniewski et al., 2006). The three SES indicators were highly correlated (r = 0.57–0.67) and loaded significantly onto one factor. The population-wide distribution of the resulting factor was divided in tertiles for analyses.

Childhood victimization

Request a detailed protocol

As previously described (Danese et al., 2017), we assessed exposure to six types of childhood victimization between birth to age 12: exposure to domestic violence between the mother and her partner, frequent bullying by peers, physical and sexual harm by an adult, and neglect.

DNA methylation data

DNA methylation was measured from Illumina 450 k Arrays in GSE55763, NAS, Framingham, and E-Risk and from Illumina EPIC 850 k Arrays in the Dunedin Study and Understanding Society. DNA was derived from whole blood samples in all studies. Dunedin Study blood draws were conducted at the cohort’s age-38 assessment during 2010–12 for the prior DunedinPoAm measure, and at the cohort’s age-45 assessment during 2017–2019 for the new DunedinPACE measure. Understanding Society blood draws were conducted in 2012. NAS blood draws were conducted during 1999–2013. Framingham blood draws were conducted during 2005–2008. E-Risk blood draws were conducted at the cohort’s age-18 assessment during 2012–13. Dunedin methylation assays were run by the Molecular Genomics Shared Resource at Duke Molecular Physiology Institute, Duke University (USA). Understanding Society and E-Risk assays were run by the Complex Disease Epigenetics Group at the University of Exeter Medical School (UK) (https://www.epigenomicslab.com). NAS methylation assays were run by the Genome Research Core of the University of Illinois at Chicago. Framingham methylation assays were run by the University of Minnesota and John’s Hopkins University (dbGaP phs000724.v9.p13). Processing protocols for the methylation data from all studies have been described previously (Dai et al., 2017; Hannon et al., 2018; Marzi et al., 2018; Panni et al., 2016).

DNA methylation clocks

Request a detailed protocol

We computed the methylation clocks proposed by Horvath, 2013, Hannum et al., 2013, Levine et al., 2018, and Lu et al., 2019 using the methylation data provided by the individual studies and the Horvath Lab’s webtool (http://dnamage.genetics.ucla.edu/new).

DunedinPACE

Request a detailed protocol

The new Dunedin Pace of Aging methylation algorithm (DunedinPACE) was developed using elastic-net regression analysis carried out in the Dunedin Study, as described in detail in the Results. The criterion variable was 20-year Pace of Aging. Development of the 20-year Pace of Aging is described in detail elsewhere (Elliott et al., 2021b). Briefly, we conducted mixed-effects growth modeling of longitudinal change in 19 biomarkers measuring integrity of the cardiovascular, metabolic, renal, hepatic, pulmonary, periodontal, and immune systems. Biomarkers were measured at the age 26, 32, 38, and 45 assessments. Biomarkers are listed in Supplementary file 1A. For each biomarker, we estimated random slopes quantifying each participant’s own rate of change in that biomarker. We then composited, the 19 slopes across the biomarkers to calculate a participant’s Pace of Aging. Pace of Aging was scaled in units representing the mean trend in the cohort, that is the average physiological change occurring during one calendar year (N = 955, M = 1, SD = 0.29). Of the N = 818 Dunedin Study members with methylation data at age 45, N = 817 had measured Pace of Aging (M = 0.99, SD = 0.30). This group formed the analysis sample to develop DunedinPACE.

To compute DunedinPACE in the GSE55763, Understanding Society, NAS, Framingham, and E-Risk datasets, we applied the scoring algorithm derived from elastic net regression in the Dunedin Study. CpG weights for the scoring algorithm are provided in Supplementary file 2. To apply the scoring algorithm, a panel of 20,000 probes that represent the underlying distribution of all probes included in the analysis to develop DunedinPACE are drawn from the target dataset. This new dataset is then quantile-normalized to match the reference distribution from the Dunedin Study. The set of 20,000 probes is composed of the 173 probes that make up the DunedinPACE algorithm and 19,827 additional probes selected at random from the full set of probes included in analysis to develop DunedinPACE i.e probes passing quality control checks, which are included on Illumina 450 k and EPIC arrays, and which were reliable at an ICC threshold of 0.4 in the analysis by Sugden et al., 2020.

Statistical analysis

Request a detailed protocol

We conducted analysis of Dunedin, GSE55763, Understanding Society, NAS, Framingham, and E-Risk, data using regression models. We analyzed continuous outcome data using linear regression. We analyzed count outcome data using Poisson and negative binomial regression. We analyzed time-to-event outcome data using Cox proportional hazard regression. To estimate intraclass correlation coefficients for technical replicates in GSE55763, we used mixed-effects regression models. For analysis of repeated-measures longitudinal DNA methylation data in the NAS, we used generalized estimating equations to account for non-independence of repeated observations of individuals (Ballinger, 2016), following the method in previous analysis of those data (Gao et al., 2018), and econometric fixed-effects regression (Wooldridge, 2012) to test within-person change over time. For analysis in E-Risk, which include data on twin siblings, we clustered standard errors at the family level to account for non-independence of data. For regression analysis, methylation measures were adjusted for batch effects by regressing the measure on batch controls and predicting residual values. Dunedin Study, Understanding Society, Framingham, and E-Risk analyses included covariate adjustment for sex (the NAS included only men). Understanding Society, NAS, and Framingham analyses included covariate adjustment for chronological age. (Dunedin and E-Risk are birth-cohort studies and participants are all the same chronological age.) Sensitivity analyses testing covariate adjustment for leukocyte distributions and smoking are reported in Supplementary file 2. Leukocyte distributions were measured from complete blood count data in the Dunedin Cohort and were estimated from DNA methylation data for other cohorts.

Code for Analysis

Request a detailed protocol

Code is available at https://github.com/danbelsky/DunedinPACE-eLife-2022 (copy archived at swh:1:rev:e38233555b70c34f84f0b73e18f1a6bc4cb0852e, Belsky, 2022a).

Code to calculate DunedinPACE from Illumina 450 k or Epic Array Data. Code is available at https://github.com/danbelsky/DunedinPACE, (copy archived at swh:1:rev:a878c3706aec2ac41bc029704cef321f7c0a4e38, Belsky, 2022b).

Appendix 1

Correlation with criterion Pace of Aging and test-retest reliability for DunedinPACE, original DunedinPoAm, and alternative versions of these algorithms developed using different sets of CpG sites.

DunedinPACE and DunedinPoAm were both developed from machine-learning analysis fitting DNA methylation data to Pace of Aging measures derived from longitudinal analysis of organ-function and blood chemistry data within the Dunedin Study.

There are two key differences between the measures. First, the Pace of Aging measure modeled to develop DunedinPoAm included data from only three repeated measures spanning 12 years of follow-up. In contrast, the Pace of Aging measure modeled to develop DunedinPACE included data from four repeated measurements collected over 20 years of follow-up. Second, DunedinPoAm was developed from a DNA methylation dataset including all probes included on both the Illumina450k and EPIC arrays. In contrast, the DNA methylation dataset used to develop DunedinPACE was restricted to a subset of CpG sites meeting a test-retest reliability threshold.

These differences should have two consequences for DunedinPACE. First, the inclusion of a 4th timepoint in Pace of Aging analysis should increase precision of measurement and result in improved fit of DNA methylation data to the 4-timepoint Pace of Aging criterion as compared to the 3-timepoint version. (A less precise Pace of Aging will be harder for the elastic net machine learning analysis to model since more of the variation will be random noise.) Second, the restriction of machine learning to more reliable CpG sites should result in improved test-retest reliability.

To further test these improvements, we conducted a sensitivity analysis. We repeated the machine-learning analysis to construct DunedinPACE using the unrestricted set of CpG sites included in the original DunedinPoAm analysis. We refer to the resulting measure as “All-probes DunedinPACE”. In parallel, we repeated the machine-learning analysis to construct DunedinPoAm using the subset of more-reliable CpG sites included in the original DunedinPACE analysis. We refer to the resulting measure as “Reliable Probes DunedinPoAm”. We then repeated our analyses testing correlation of the DNA methylation measures with the criterion Pace of Aging and our analysis of test-retest reliability.

Consistent with the expectation that a 4th timepoint would increase precision of Pace of Aging measurement and result in improved criterion fit, both versions of DunedinPACE showed stronger correlation with 4-timepoint Pace of Aging than did the versions of DunedinPoAm with 3-timepoint Pace of Aging. For algorithms trained on all probes, r = 0.87 DunedinPACE vs. r = 0.58 for original DunedinPoAm. For algorithms trained on reliable probes, r = 0.77 for DunedinPACE vs. r = 0.65 for DunedinPoAm.

Consistent with the expectation that training the algorithm using more reliable probes would improve test-retest reliability, the reliable-probes-trained DunedinPACE showed superior test-retest reliability to the all-probes trained DunedinPACE in the test-retest reliability dataset published by Lehne et al. (GSE55763) (reliable-probes ICC = 0.96 vs 0.72 for all-probes). ICCs were similar for the two version of DunedinPoAm (reliable-probes ICC = 0.90 vs 0.89 for all probes).

Data availability

Datasets are available from the data owners. Data from the Dunedin and E-Risk Study can be accessed through agreement with the Study investigators. Instructions are available at https://sites.google.com/site/moffittcaspiprojects/. The data access application form can be downloaded here: https://sites.google.com/site/moffittcaspiprojects/forms-for-new-projects/concept-paper-template. Data from the Understanding Society Study is available through METADAC at https://www.metadac.ac.uk/ukhls/. All details are on the Metadac website (https://www.metadac.ac.uk/data-access-through-metadac/). The data access application form can be found here https://www.metadac.ac.uk/files/2019/02/v2.41-UKHLS-METADAC-application-form-2019-2hak8bv.docx. Data from the Normative Aging Study were obtained from the Study investigators. Data are accessible through dbGaP, accession phs000853.v1.p1. Data from the Framingham Heart Study were obtained from dbGaP, accession phs000007.v32.p13. GSE55763 is a publicly available dataset available from the Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE55763).

The following previously published data sets were used
    1. Lehne
    (2021) NCBI Gene Expression Omnibus
    ID GSE55763. A coherent approach for analysis of the Illumina HumanMethylation450 BeadChip improves data quality and performance in epigenome-wide association studies.

References

    1. Corcoran DL
    2. Sugden K
    3. Caspi A
    4. Belsky DW
    (2022)
    pipeline_for_Illumina450k_and_EPIC_array_data
    DunedinPACE 1:a878c37.
    1. Moffitt TE
    2. Team E
    (2002) Teen-aged mothers in contemporary Britain
    Journal of Child Psychology and Psychiatry, and Allied Disciplines 43:727–742.
    https://doi.org/10.1111/1469-7610.00082
  1. Book
    1. Wooldridge JM
    (2012)
    Introductory Econometrics: A Modern Approach (5th ed)
    Cengage Learning.
  2. Conference
    1. Yanchun B
    (2017) Understanding Society: DNA Methylation Data Based on Illumina Methylation EPIC Array
    University of Essex Institute for Social and Economic Research, NatCen Social Research, and University of Exeter Medical School.
    https://doi.org/10.1016/j.envres.2020.109971

Decision letter

  1. Joris Deelen
    Reviewing Editor; Max Planck Institute for Biology of Ageing, Cologne, Germany
  2. Jessica K Tyler
    Senior Editor; Weill Cornell Medicine, Cornell University, Ithaca, United States
  3. Matthew Suderman
    Reviewer; University of Bristol, United Kingdom
  4. Joris Deelen
    Reviewer; Max Planck Institute for Biology of Ageing, Germany

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Decision letter after peer review:

Thank you for submitting your article "Quantification of the pace of biological aging in humans through a blood test: the DunedinPACE DNA methylation algorithm" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, including Joris Deelen as the Reviewing Editor and Reviewer #3, and the evaluation has been overseen by Jessica Tyler as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Matthew Suderman (Reviewer #2).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

1. The reviewers all agree that it is unclear what the added value of DunedinPACE is over the previously generated DundinPOAM and, even more, the widely used GrimAge methylation biomarkers. This should be discussed more thoroughly and will require some additional analyses (see individual review reports).

2. The discussion between the reviewers also brought forward that the work by Levine and colleagues (doi: https://doi.org/10.1101/2021.04.16.440205), although shortly mentioned in the introduction, has not sufficiently been taken into account by the authors. This preprint has several implications for the current manuscript. First, the currently used methylation biomarkers, including GrimAge, will soon be updated using the correction proposed by Levine et al. and will outperform the versions used in the current manuscript. Thus, the iterative improvement in DunedinPACE vs DunedinPoAm may be the only relevant comparison. Second, the current manuscript uses test-retest of individual CpG probes to refine DunedinPACE. However Levine et al. show strong evidence that this strategy may improve performance for a single clock (e.g. better performance in DunedinPoAM to DunedinPACE) but that this may not generalize to other clocks. Rather, technical variability was best improved by PCA to extract a 'shared signal' vs individual CpGs. At minimum, the authors should consider various approaches to reduce technical reliability rather than focus solely on individual CpGs, including creating a methylation biomarker using principal components.

The authors should also address the additional points mentioned in the individual review reports below.

Reviewer #1:

This study details development of an improved iteration of a DNA methylation biomarker based on Pace of Aging measures developed in healthy, primarily Caucasian adults up to age 45 years in the highly unique longitudinal Dunedin Study. The new methylation biomarker is DunedinPACE; the predecessor DunedinPoAm was reported in Belsky et al. 2020 eLife 9,354870.

Strengths

The key contribution of this study is to address limitations in previous methylation biomarkers: (1) Pace of Aging measures now encompass a 20-year time span vs 12 years in Dunedin study; (2) training population includes adults up to 45 years (vs 38 years); and (3) methylation model is restricted only to probes that meet a minimum threshold for test re-test reliability.

1. DunedinPACE has good test-retest reliability, which is highly relevant to biomarker performance in clinical studies and interventions testing.

2. Not unexpectedly, the test-retest reliability is stronger for the new DunedinPACE vs. earlier DunedinPoAm, possibly owing to introduction of minimum probe reliability thresholds, though this is not explicitly tested.

3. The analyses presented generally follow established tests of criterion and construct validity across a set of cohort studies previously used to develop its predecessor DundedinPoAm (see Belsky et al. 2020 eLife 9,354870). The parallel strategy makes the seemingly complex set of validation studies a relatively straightforward comparison across biomarker iterations.

4. The effect sizes in cross-cohort validation analyses are larger for DunedinPACE relative to the foundational DunedinPoAm, and the effect sizes meet (GrimAge) or surpass effect sizes for other existing DNAm biomarkers. This supports strength of the new PACE biomarker iteration relative to most others.

5. Conclusions are generally well supported by data and authors refrain from extraneous discussion.

Weaknesses

1. Though several key limitations in DunedinPACE's predecessors were addressed, it is unfortunate that readers are not provided details regarding their relative contributions to overall biomarker performance.

2. It would be highly informative for future biomarker development to know if and to what extent population characteristics (cohort, age, or duration of observations for Pace of Aging measure) drive improved performance of DunedinPACE vs. DunedinPoAm relative to test-retest reliability or duration of longitudinal observation.

3. DunedinPACE performed similarly (reported effect sizes) to GrimAge, yet this equivalence is not addressed in discussion or conclusion. It is unclear under which conditions should one biomarker be prioritized over another if effect sizes are similar.

4. Forced language regarding DunedinPACE as a proxy for Pace of Aging muddles results presentation and section headings; this may inadvertently misrepresent select analyses.

5. Remaining concerns and suggestions are to improve readability, clarify nomenclature, and strengthen presentation of results, and do not influence primary conclusions or implications.

Title:

6. The title is regurgitated from the eLife 2020 publication but with a modification after the colon. This tedious for query building, search engines, and reference lists. The authors should thus consider adapting it.

Contribution of individual iterative changes:

7. To address individual contributions of each of the prior limitations, the authors could consider supporting analyses to address:

• Ddoes the inclusion of persons who are 45 years and now exhibiting detectable declines in health / function improve performance of the methylation biomarker, or is the time span (12 vs 20 years) the key determinant for calibration for improved performance?

– Potential approach: restrict model to population aged 32-45 years (~12 years, but older) and compare performance to current DunedinPACE and previous DunedinPoAm.

• Is minimum reliability threshold the key to improved performance or characteristics of the population used to calibrate the methylation biomarker?

– Potential approach: (a) revisit DunedinPoAm – restrict the existing biomarker to probes meeting reliability threshold, or (b) release test-retest reliability restrictions for probes included in DunedinPACE and evaluate.

Implications:

8. DunedinPACE is an iterative change to DunedinPoAm; clearly showing how each of these changes impact the resultant model bolsters significance.

9. DunedinPACE is tested head-to-head with previous derivations of DNAm clocks, but without insight regarding if or how changes leading to DunedinPACE could be introduced to rebuild better a GrimAge, PhenoAge, etc.

10. If primary determinant of improved performance is including aged >45 years, what does this mean for generalizability of clocks tuned in young/healthy population to trials in older adults or those in poorer health.

11. If analyses show that duration of longitudinal observation is key (20 years vs 12 years), then would future Dunedin derivations be expected to consistently outperform the previous versions and what would this mean for durability of the DunedinPACE as a biomarker?

GrimAge:

12. Addressing GrimAge vs DunedinPACE similarity in effect sizes may require some speculation, but would aid reader interpretation of results and could be simply handled with a short sentence or two in discussion.

Language – DunedinPACE as Pace of Aging Proxy:

13. Language referring to DunedinPACE and Pace of Aging is sometimes confusing, which could be attributed to the forced representation of DunedinPACE as a proxy for Pace of Aging.

• This results in instances of merged language (pg 12 "DunedinPACE Pace of Aging").

• The Results section and headings also suffer from forced association (marketing over reporting findings?). Examples:

– (Results pg 9) "DunedinPACE is indicates faster Pace of Aging in Chronologically Older Individuals." Results in text and Figure 3A show association with chronological age.

– (Results pg 10) "DunedinPACE shows faster Pace of Aging in individuals measured to be…" Forcing reference to Pace of Aging makes the heading cumbersome ("DunedinPACE is associated with other epigenetic clocks, measures of biologic age, and self-rated health").

– Authors are recommended to use simplicity and clarity throughout results and save comment about DunedinPACE as possible a proxy for Pace of Aging for the discussion.

Results, Text and Figure 3 Alignment:

– Page 9-10. "DunedinPACE indicates faster Pace of Aging in Chronologically Older.."

– The text opens with a sentences about mortality risk, but the only figure referenced is association with chronological age in Understanding Society study (Figure 3 Panel A). Recommend to omit or move to mortality risk references to Results section on NAS.

– The second paragraph provides information about exposure histories across birth cohorts. These results distract from the flow of results reporting in Figure 3 and subsequent section. While within-individual change may be supportive they are not essential and detract from readability.

For consideration (not mandatory): Naming Conventions

14. One improvement of DunedinPACE is the name; it is much easier to read / say compared with DunedinPoAm. However, would "PACE" or "PACEm" be sufficient? Adding "Dunedin…" seems to suggest that DunedinPACE or DunedinPoAm are specific to the foundational study of origin compared with GrimAge or PhenoAge.

15. "Faster" and "Slower" are used repeatedly to refer to higher vs. lower DunedinPACE (single-timepoint) compared to Pace of Aging (time component included). Using "faster" or "slower" may be appropriate for a Pace-proxy unlike most clocks. But referring to change in single timepoint DunedinPACE as "faster" is confusing, even if it is proposed as a proxy for Pace of Aging

Reviewer #2:

Although the paper is well-written, findings are not presented in a way that allows the reader to answer the most important question of the paper: how does DunedinPACE compare to DunedinPOAM? Little can be concluded from the improved performance in the Dunedin study data from which DunedinPACE and DunedinPOAM were both derived. In independent datasets, performance appears to be almost identical. Test-retest evaluations in the Sugden dataset similarly are difficult to interpret because probe reliability was calculated using the Sugden dataset.

In all evaluations using data used to derive DunedinPACE, authors should clearly state the potential for inflated performance. This is particularly true for comparisons with DundinPOAM which was trained on a subset of DunedinPACE data and previously published aging estimates Horvath/Hannum/Phenoage/Grimage trained entirely on external data.

Reviewer #3:

The manuscript by Belsky and colleagues reports the development of an improved methylation biomarker based on longitudinal health data (Pace of Aging) from the Dunedin Study. The authors have now used an additional wave of measurements and excluded methylation probes with low reliability based on recent publications. They subsequently associated their improved biomarker, which they called DunedinPACE, with many different health-related outcomes, such as physical functioning, morbidity and mortality. They show that DunedinPACE shows comparable results as GrimAge, a methylation biomarker based on mortality.

The major strength of the study is that the authors have managed to improve their previous methylation biomarker (DunedinPoAm) by adding additional data. They also show that DunedinPACE does reasonable well when associated with health-related outcomes. However, it is unclear what the added value of DunedinPACE is in comparison to GrimAge, because it does not outperform this previous biomarker in its association with any of the outcomes. Moreover, the provided data on the most relevant health-related outcomes, i.e. morbidity and mortality, is limited and it is unclear if the methylation biomarker would outperform other omics-based biomarkers (i.e. metabolomics and proteomics).

– The authors should highlight the added value of DunedinPACE in comparison to GrimAge. It looks like they do equally well in the analysis they performed, so what would be the advantage of using DunedinPACE over GrimAge, given that the latter is already widely used.

– The authors should also look at cause-specific mortality and morbidity to see if DunedinPACE reflects general health or is specific to one major disease, e.g. cardiovascular disease.

– The authors should test the predictive ability of DunedinPACE, ideally in comparison to currently used clinical biomarkers and biomarkers based on other omics data (see PMID: 34145379 and PMID: 31431621). If a direct comparison is not possible, the existence of these non-methylation biomarkers should still be mentioned, including their effect sizes for morbidity/mortality (if available).

– The authors should test the association of DunedinPACE with morbidity and mortality in different age-groups (i.e. stratify their data) to make sure the associations are not driven by early-life morbidity and mortality.

– It would be interesting if the authors can provide a biological interpretation (e.g. functional annotation / genetic association analysis) of the CpGs that are included in the DunedinPACE biomarker, as already done for some of the other (methylation-based) biomarkers (see for example PMID: 34187551 and PMID: 34038024).

https://doi.org/10.7554/eLife.73420.sa1

Author response

Essential revisions:

1. The reviewers all agree that it is unclear what the added value of DunedinPACE is over the previously generated DundinPOAM and, even more, the widely used GrimAge methylation biomarkers. This should be discussed more thoroughly and will require some additional analyses (see individual review reports).

Thank you for the opportunity to clarify value added of DunedinPACE. PoAm was the first iteration and PACE is an advance on it. Both were derived in the Dunedin Longitudinal Study. The value added of DunedinPACE over PoAm is primarily technical and consists of two key advances: (1) PACE offers improved precision of DNA methylation measurement of Pace of Aging; DunedinPACE is nearly twice as accurate as DunedinPoAm in predicting longitudinal Pace of Aging (correlations of r=0.6 and r=0.8 translate into coefficients of determination of R2=0.36 and 0.64). (2) PACE offers improved technical reliability of the DNA methylation measurement (ICC>.95 in technical replicates). High reliability is essential for detection of change over time in RCTs. These two advances are reviewed in the first and second paragraphs of the discussion.

The value added of DunedinPACE over DNAmethylation clocks is in its design. Many features of a person’s biology may predict their risk of disease and death. The 2nd-generation PhenoAge and GrimAge clocks operationalize these. But only some of these features are aging. DunedinPACE operationalizes aging itself as decline in physiological integrity over years. DunedinPACE has three design features that isolate aging signal:

(1) The Dunedin cohort were all born in the same year. This eliminates bias from historical-exposure effects, including differences between people born in different years in their mothers’ health and environment when they were in-utero, their childhood nutrition, their exposure to childhood infections, and their exposure to environmental toxicants, such as air pollution from leaded gasoline and second-hand tobacco smoke in the workplace.

(2) Dunedin cohort members were young adults at the time of baseline measurements. This eliminates survival bias; none of the cohort had died of aging-related causes by the time of baseline Pace of Aging assessment. It also separates aging from age-related disease; physiological changes observed beginning in cohort members’ 20s represent pre-clinical changes in system integrity that are causes of disease, not disease sequelae.

(3) DunedinPACE is modeled on decline in the integrity of multiple organ systems during adult life. This isolates aging-related changes from deficits established early in development.

In contrast, clocks were developed on mixed-age populations and cannot distinguish aging from historical-exposure effects. Clocks were developed from mixed-age or older populations, introducing survival bias and uncertainty about what is aging versus disease. Finally, clocks are based on a single cross-sectional measurement and do not distinguish aging-related changes from deficits established early in development. We have added a final paragraph to the introduction to clarify these points.

“As a new DNA methylation measure of biological aging, DunedinPACE joins a well-established battery of DNA methylation measures known as “clocks”, so-named for their accurate prediction of chronological age (Horvath and Raj, 2018). […] DunedinPACE, therefore, represents a complementary tool for DNA methylation quantification of biological aging.”

Consistent with these distinguishing features, DunedinPACE captures information about aging that is not measured by the clocks. We have added a series of multivariate analyses in the Framingham Heart Study cohort (which was used to develop GrimAge) that show DunedinPACE predicts mortality, incident cardiovascular disease, and incident disability over and above the clocks, including GrimAge. This is a rigorous test since GrimAge was developed using the Framingham data. Results are reported in the final paragraph of the Results section and detailed in Supplementary file 1C.

“Incremental predictive value of DunedinPACE in an analysis of morbidity, disability, and mortality. DunedinPACE is distinct from DNA methylation clocks in the method of its development and in its interpretation. […] Results for all models are reported in Supplementary file 1C.”

2. The discussion between the reviewers also brought forward that the work by Levine and colleagues (doi: https://doi.org/10.1101/2021.04.16.440205), although shortly mentioned in the introduction, has not sufficiently been taken into account by the authors. This preprint has several implications for the current manuscript. First, the currently used methylation biomarkers, including GrimAge, will soon be updated using the correction proposed by Levine et al. and will outperform the versions used in the current manuscript. Thus, the iterative improvement in DunedinPACE vs DunedinPoAm may be the only relevant comparison. Second, the current manuscript uses test-retest of individual CpG probes to refine DunedinPACE. However Levine et al. show strong evidence that this strategy may improve performance for a single clock (e.g. better performance in DunedinPoAM to DunedinPACE) but that this may not generalize to other clocks. Rather, technical variability was best improved by PCA to extract a 'shared signal' vs individual CpGs. At minimum, the authors should consider various approaches to reduce technical reliability rather than focus solely on individual CpGs, including creating a methylation biomarker using principal components.

Thank you for this suggestion. First, as our paper reports, DunedinPACE already demonstrates excellent test-retest reliability (ICC=0.96-0.97). So, the potential for the approach proposed in the Higgins-Chen et al. preprint to improve the test-retest reliability of our measure is virtually nil. Moreover, the method introduced by Higgins-Chen was designed to address the limitation that, when DNAm clocks were re-trained using subsets of more reliable probes, the resulting clock measures were less predictive of morbidity and mortality as compared to their original versions. As our data show, that is not the case for the reliable-probes-based DunedinPACE; DunedinPACE effect-sizes are the same as or larger than effect-sizes for the original DunedinPoAm, which did not use reliable probes. Thus, the Higgins-Chen et al. approach may be a useful a solution to a thorny problem that uniquely afflicts clocks such as PhenoAge, but that problem does not afflict DunedinPACE and the corrective approach is not needed by DunedinPACE.

However, we agree that it is important to acknowledge that the analysis by Higgins-Chen et al. suggests that the approach we used does not generalize to existing epigenetic clocks. Therefore, in addition to the text in our article which already acknowledges the potential for methods such as those proposed in the Higgins-Chen et al. preprint to improve reliabilities of existing clocks, we now also acknowledge that the method we used may not generalize to the epigenetic clocks.

“Comparison of DunedinPACE reliability to reliabilities for the original DunedinPoAm and the DNA methylation clocks are shown in Supplemental Figure 2. DunedinPACE is as reliable or more so as compared to the GrimAge clock and more reliable in comparison to the other clocks, although methods have been proposed that may improve reliabilities for the other clocks (Higgins-Chen et al., 2021).” (results p10 para 4)

“Interestingly, in contrast to reports for some epigenetic clocks (Higgins-Chen et al., 2021), the restriction of machine learning analysis to a more-reliably-measured subset of CpG sites on the Illumina EPIC array did not harm criterion validity of DunedinPACE; effect-sizes for prediction of morbidity and mortality were as large or larger in comparison to DunedinPoAm.” (discussion p17 para 2)

The authors should also address the additional points mentioned in the individual review reports below.

Reviewer #1:

This study details development of an improved iteration of a DNA methylation biomarker based on Pace of Aging measures developed in healthy, primarily Caucasian adults up to age 45 years in the highly unique longitudinal Dunedin Study. The new methylation biomarker is DunedinPACE; the predecessor DunedinPoAm was reported in Belsky et al. 2020 eLife 9,354870.

Strengths

The key contribution of this study is to address limitations in previous methylation biomarkers: (1) Pace of Aging measures now encompass a 20-year time span vs 12 years in Dunedin study; (2) training population includes adults up to 45 years (vs 38 years); and (3) methylation model is restricted only to probes that meet a minimum threshold for test re-test reliability.

1. DunedinPACE has good test-retest reliability, which is highly relevant to biomarker performance in clinical studies and interventions testing.

2. Not unexpectedly, the test-retest reliability is stronger for the new DunedinPACE vs. earlier DunedinPoAm, possibly owing to introduction of minimum probe reliability thresholds, though this is not explicitly tested.

3. The analyses presented generally follow established tests of criterion and construct validity across a set of cohort studies previously used to develop its predecessor DundedinPoAm (see Belsky et al. 2020 eLife 9,354870). The parallel strategy makes the seemingly complex set of validation studies a relatively straightforward comparison across biomarker iterations.

4. The effect sizes in cross-cohort validation analyses are larger for DunedinPACE relative to the foundational DunedinPoAm, and the effect sizes meet (GrimAge) or surpass effect sizes for other existing DNAm biomarkers. This supports strength of the new PACE biomarker iteration relative to most others.

5. Conclusions are generally well supported by data and authors refrain from extraneous discussion.

Weaknesses

1. Though several key limitations in DunedinPACE's predecessors were addressed, it is unfortunate that readers are not provided details regarding their relative contributions to overall biomarker performance.

2. It would be highly informative for future biomarker development to know if and to what extent population characteristics (cohort, age, or duration of observations for Pace of Aging measure) drive improved performance of DunedinPACE vs. DunedinPoAm relative to test-retest reliability or duration of longitudinal observation.

Thank you for these suggestions. We have conducted additional analyses to refine inference about which aspects of DunedinPACE development contributed to the key improvements over DunedinPoAm, i.e. improved fit to the Pace of Aging criterion and improved test-retest reliability. Specifically, we created two new algorithms: (1) a new version of DunedinPACE trained to fit 20-year Pace of Aging from all probes shared on the EPIC and 450k arrays (referred to as “all probes DunedinPACE”); and (2) a version of DunedinPoAm trained from the subset of probes meeting the reliability threshold set by Sugden. (referred to as “reliable probes DunedinPoAm”). Analysis of criterion fit to Pace of Aging and test-retest reliability are reported in Supplemental Results and mentioned in the Results section (p11, para 1). There are two key findings from this analysis: First, all versions trained on 20-year, 4-time-point Pace of Aging show higher correlation with their Pace of Aging criterion as compared to all versions trained on 12-year, 3-time-point Pace of Aging. This result indicates that the improved criterion fit of DuneindPACE comes from inclusion of additional follow-up in the Pace of Aging measure. Second, all algorithms trained on the subset of reliable probes showed higher test-retest reliability as compared with all algorithms trained in the complete set of probes. This result indicates that the improved test-retest reliability of DunedinPACE comes from the restriction of elastic net regression analysis to the subset of reliable probes.

3. DunedinPACE performed similarly (reported effect sizes) to GrimAge, yet this equivalence is not addressed in discussion or conclusion. It is unclear under which conditions should one biomarker be prioritized over another if effect sizes are similar.

Thank you for the opportunity to clarify. We view DunedinPACE as a measure complementary to GrimAge. As explained in our response to Essential Revisions 1, DunedinPACE is distinct from GrimAge in its design and its meaning. In addition, it also captures unique information about risk for disease and death; i.e. DunedinPACE improves prediction of morbidity and mortality over and above GrimAge. This is detailed in the final paragraph of the Results section and in Supplementary file 1C.

4. Forced language regarding DunedinPACE as a proxy for Pace of Aging muddles results presentation and section headings; this may inadvertently misrepresent select analyses.

Done. Thank you for bringing this to our attention. The text has been revised for clarity.

5. Remaining concerns and suggestions are to improve readability, clarify nomenclature, and strengthen presentation of results, and do not influence primary conclusions or implications.

Title:

6. The title is regurgitated from the eLife 2020 publication but with a modification after the colon. This tedious for query building, search engines, and reference lists. The authors should thus consider adapting it.

Done. The new title is “DunedinPACE: A DNA methylation biomarker of the Pace of Aging”.

Contribution of individual iterative changes:

7. To address individual contributions of each of the prior limitations, the authors could consider supporting analyses to address:

• Does the inclusion of persons who are 45 years and now exhibiting detectable declines in health / function improve performance of the methylation biomarker, or is the time span (12 vs 20 years) the key determinant for calibration for improved performance?

– Potential approach: restrict model to population aged 32-45 years (~12 years, but older) and compare performance to current DunedinPACE and previous DunedinPoAm.

• Is minimum reliability threshold the key to improved performance or characteristics of the population used to calibrate the methylation biomarker?

– Potential approach: (a) revisit DunedinPoAm – restrict the existing biomarker to probes meeting reliability threshold, or (b) release test-retest reliability restrictions for probes included in DunedinPACE and evaluate.

Implications:

8. DunedinPACE is an iterative change to DunedinPoAm; clearly showing how each of these changes impact the resultant model bolsters significance.

9. DunedinPACE is tested head-to-head with previous derivations of DNAm clocks, but without insight regarding if or how changes leading to DunedinPACE could be introduced to rebuild better a GrimAge, PhenoAge, etc.

This is an interesting idea. But, lacking the data used to train the other clocks, we are unable to conduct the same analysis. We now note that the Higgins-Chen preprint explored restriction of CpG sites based on ICC criteria for good test-retest reliability and found this approach was generally harmful; it reduced the criterion validity of their clocks. However restriction of CpG sites to reliable sites did not reduce criterion validity for DunedinPACE (Discussion p17 para 2, text quoted above in response to Essential Revision 2).

10. If primary determinant of improved performance is including aged >45 years, what does this mean for generalizability of clocks tuned in young/healthy population to trials in older adults or those in poorer health.

Thank you for the opportunity to clarify. Prediction of morbidity and mortality in older adults in the Normative Aging Study and Framingham cohorts in our paper is similar for DunedinPoAm (developed from Pace of Aging measured over ages 26-38) as for DunedinPACE (developed from Pace of Aging measured over ages 26-45). Therefore, the age range used in measuring Pace of Aging does not seem to be a critical feature of prediction of morbidity and mortality in older adults. DunedinPACE, though trained on data from age 26 to 45, predicts morbidity and mortality in cohorts in their 70s and 80s.

11. If analyses show that duration of longitudinal observation is key (20 years vs 12 years), then would future Dunedin derivations be expected to consistently outperform the previous versions and what would this mean for durability of the DunedinPACE as a biomarker?

We expect that, when another wave of data from the Dunedin Study is collected in 5-7 years, a new Pace of Aging measurement will be evaluated. At that time, we will be able to clarify whether longer follow-up intervals extending into later stages of life contribute new and important information to a Pace of Aging DNA methylation measure. Until then, geroscience needs the best measurement tools available to evaluate intervention effects today. DunedinPACE is an improvement over what is already available.

GrimAge:

12. Addressing GrimAge vs DunedinPACE similarity in effect sizes may require some speculation, but would aid reader interpretation of results and could be simply handled with a short sentence or two in discussion.

Thank you for this suggestion. We have added text to the introduction (p7-8), results (p15-16), and discussion (p17-18). These changes are also summarized in our response to Essential Revision 1 above.

Language – DunedinPACE as Pace of Aging Proxy:

13. Language referring to DunedinPACE and Pace of Aging is sometimes confusing, which could be attributed to the forced representation of DunedinPACE as a proxy for Pace of Aging.

• This results in instances of merged language (pg 12 "DunedinPACE Pace of Aging").

• The Results section and headings also suffer from forced association (marketing over reporting findings?). Examples:

– (Results pg 9) "DunedinPACE is indicates faster Pace of Aging in Chronologically Older Individuals." Results in text and Figure 3A show association with chronological age.

– (Results pg 10) "DunedinPACE shows faster Pace of Aging in individuals measured to be…" Forcing reference to Pace of Aging makes the heading cumbersome ("DunedinPACE is associated with other epigenetic clocks, measures of biologic age, and self-rated health").

– Authors are recommended to use simplicity and clarity throughout results and save comment about DunedinPACE as possible a proxy for Pace of Aging for the discussion.

Done. We have revised text and figure titles.

Results, Text and Figure 3 Alignment:

– Page 9-10. "DunedinPACE indicates faster Pace of Aging in Chronologically Older.."

– The text opens with a sentences about mortality risk, but the only figure referenced is association with chronological age in Understanding Society study (Figure 3 Panel A). Recommend to omit or move to mortality risk references to Results section on NAS.

Thank you for the opportunity to clarify. The text referencing mortality risk explains why we anticipate a correlation between DunedinPACE and chronological age. We have further revised the text for clarity.

“In demography, the pace of biological aging in a population can be estimated from the rate of increase in mortality risk from younger to older chronological ages (Hägg et al., 2019). In humans and many other species, the increase in mortality risk with advancing chronological age is nonlinear, suggesting that the pace of biological aging accelerates as we grow older (Gompertz, 1825; Kirkwood, 2015; Olshansky and Carnes, 1997). We tested if, consistent with this hypothesis, DunedinPACE was faster in chronologically older as compared to younger individuals in data from the Understanding Society Study (n=1175, age range 28-95). Chronologically older Understanding Society participants had faster DunedinPACE as compared to younger ones (r=0.32, Figure 3 Panel A). This correlation was threefold larger as compared to the correlation for the original DunedinPoAm (r=0.11). ” (p11 para 3)

– The second paragraph provides information about exposure histories across birth cohorts. These results distract from the flow of results reporting in Figure 3 and subsequent section. While within-individual change may be supportive they are not essential and detract from readability.

Thank you for this suggestion. We think these data on within-individual change are essential for establishing the independence of DunedinPACE correlations with chronological age from historical exposures that vary by birth year and may alter the methylome (i.e. “cohort effects”). While cohort effects are not yet widely appreciated as a source of bias in biology, they are well established in human epidemiologic research. Therefore, we have retained the text.

For consideration (not mandatory): Naming Conventions

14. One improvement of DunedinPACE is the name; it is much easier to read / say compared with DunedinPoAm. However, would "PACE" or "PACEm" be sufficient? Adding "Dunedin…" seems to suggest that DunedinPACE or DunedinPoAm are specific to the foundational study of origin compared with GrimAge or PhenoAge.

Thank you for this suggestion. We appreciate the inclusion of the name of the study in the name of our measure is somewhat cumbersome. But we also think it is essential. Data sources used in the derivation of machine-learning-based measures can affect the measure in many ways – something we acknowledge in the limitations section of our discussion. In our view, we should all think of the Framingham-GrimAge, the InCHIANTI-PhenoAge, etc. We have therefore elected to retain our original study name to inform the science. In addition, there is an ethical reason. Dunedin Study participants were consulted in focus groups about their views on using Study data to derive a measure that might have intellectual property, and they asked that their contribution of repeated blood samples be acknowledged and recognized in the naming of the measure. Given the current unpleasant situation with the family of Henrietta Lacks in reaction to commercial use of HELA cells, consulting the Study participants was good policy.

15. "Faster" and "Slower" are used repeatedly to refer to higher vs. lower DunedinPACE (single-timepoint) compared to Pace of Aging (time component included). Using "faster" or "slower" may be appropriate for a Pace-proxy unlike most clocks. But referring to change in single timepoint DunedinPACE as "faster" is confusing, even if it is proposed as a proxy for Pace of Aging.

Thank you for this suggestion. We have retained the language of faster/slower as DunedinPACE is not a clock and does not measure an “age” value. DunedinPACE is a prediction of Pace of Aging. Higher values correspond to faster Pace of Aging and lower values to slower Pace of Aging. Faster/slower therefore aligns with the measure. In addition, this precise language helps to distinguish DunedinPACE from results for the clocks, which were designed to generate values younger/older, but not faster/slower.

Reviewer #2:

Although the paper is well-written, findings are not presented in a way that allows the reader to answer the most important question of the paper: how does DunedinPACE compare to DunedinPOAM? Little can be concluded from the improved performance in the Dunedin study data from which DunedinPACE and DunedinPOAM were both derived. In independent datasets, performance appears to be almost identical. Test-retest evaluations in the Sugden dataset similarly are difficult to interpret because probe reliability was calculated using the Sugden dataset.

We report head-to-head comparisons of DunedinPACE with DunedinPoAm in the text and in the supplemental figures and specifically note the outcomes of these comparisons in the Results section. We have highlighted these comparisons in the revised submission.

Regarding Test-retest reliability: the probe reliability threshold used to define the CpG set used to develop DunedinPACE was defined in only one of the datasets used to evaluate test-retest reliability (Sugden B n=350 450k-EPIC, Supplemental Figure 2). Results are consistent across all datasets, arguing against the finding being an artifact of the Sugden dataset.

In all evaluations using data used to derive DunedinPACE, authors should clearly state the potential for inflated performance. This is particularly true for comparisons with DundinPOAM which was trained on a subset of DunedinPACE data and previously published aging estimates Horvath/Hannum/Phenoage/Grimage trained entirely on external data.

Done. This issue applies to the Dunedin data, but not to the several other datasets analyzed. We now note that the analysis of physical and cognitive functioning and subjective signs of aging within the Dunedin cohort may result in some upward bias to effect-size estimates for DunedinPACE and DunedinPoAm within that cohort.

“DunedinPACE and DunedinPoAm were developed from analysis of Pace of Aging in the Dunedin cohort; effect-sizes for these measures may be over-estimated within this cohort.” (p10 para 2)

Reviewer #3:

The manuscript by Belsky and colleagues reports the development of an improved methylation biomarker based on longitudinal health data (Pace of Aging) from the Dunedin Study. The authors have now used an additional wave of measurements and excluded methylation probes with low reliability based on recent publications. They subsequently associated their improved biomarker, which they called DunedinPACE, with many different health-related outcomes, such as physical functioning, morbidity and mortality. They show that DunedinPACE shows comparable results as GrimAge, a methylation biomarker based on mortality.

The major strength of the study is that the authors have managed to improve their previous methylation biomarker (DunedinPoAm) by adding additional data. They also show that DunedinPACE does reasonable well when associated with health-related outcomes. However, it is unclear what the added value of DunedinPACE is in comparison to GrimAge, because it does not outperform this previous biomarker in its association with any of the outcomes. Moreover, the provided data on the most relevant health-related outcomes, i.e. morbidity and mortality, is limited and it is unclear if the methylation biomarker would outperform other omics-based biomarkers (i.e. metabolomics and proteomics).

– The authors should highlight the added value of DunedinPACE in comparison to GrimAge. It looks like they do equally well in the analysis they performed, so what would be the advantage of using DunedinPACE over GrimAge, given that the latter is already widely used.

Thank you for this suggestion. As explained in response to Essential Revision 1 and in a new paragraph added to the introduction (p7-8), DunedinPACE represents distinct information about aging from GrimAge. Moreover, empirically, DuedinPACE is an independent predictor of morbidity, disability, and mortality (Results p15-16; Supplemental Table 5). Therefore, DunedinPACE is complementary to GrimAge (Discussion p18 para 2).

– The authors should also look at cause-specific mortality and morbidity to see if DunedinPACE reflects general health or is specific to one major disease, e.g. cardiovascular disease.

Thank you for this suggestion. We have added time-to-event analysis of cardiovascular disease and stroke and repeated-measures analysis of incident disability in the Framingham Heart Study Cohort. See results p14 para 1, Supplemental Table 4, and Supplemental Figure 4 panels C and D.

We did not include analysis of cause-specific mortality because the number of deaths from individual causes is too small in the datasets available to us for confident ascertainment of differential associations with different causes of death. We now acknowledge this as a limitation (Discussion p18 para 2).

In response to the reviewer’s request, we did conduct the analysis using the data available from the Framingham Heart Study (which classifies deaths as due to Coronary Heart Disease, other Cardiovascular Disease, and all other causes). The effect-size for cardiovascular disease is HR=1.46 95% CI [1.23-1.72]. The effect-size for non-CVD causes is HR=1.70 95% CI [1.55-1.87].

– The authors should test the predictive ability of DunedinPACE, ideally in comparison to currently used clinical biomarkers and biomarkers based on other omics data (see PMID: 34145379 and PMID: 31431621). If a direct comparison is not possible, the existence of these non-methylation biomarkers should still be mentioned, including their effect sizes for morbidity/mortality (if available).

Thank you for this suggestion. We now acknowledge the exciting progress in proteomics and metabolomics research on aging in the limitations section of the discussion:

“Recent reports suggest valuable information about aging and healthspan may also be captured in metabolomic and proteomic datasets (Deelen et al., 2019; Eiriksdottir et al., 2021; Jansen et al., 2021; Lehallier et al., 2019; Tanaka et al., 2020), although such datasets are as yet rare. Future studies should compare DunedinPACE to measures derived from these biological levels of analysis.” (p18 para 2)

– The authors should test the association of DunedinPACE with morbidity and mortality in different age-groups (i.e. stratify their data) to make sure the associations are not driven by early-life morbidity and mortality.

Thank you for the opportunity to clarify. All mortality analyses were conducted using data from older adults (Normative Aging Study men were aged 77 years on average at DNA methylation baseline; Framingham participants were aged 66 years on average a DNA methylation baseline; see results p13-14). Therefore, early-life morbidity and mortality cannot explain the results observed.

– It would be interesting if the authors can provide a biological interpretation (e.g. functional annotation / genetic association analysis) of the CpGs that are included in the DunedinPACE biomarker, as already done for some of the other (methylation-based) biomarkers (see for example PMID: 34187551 and PMID: 34038024).

We have conducted the analysis suggested by the reviewer and results are included below. However, it is our preference not to include these in the manuscript. The reason is that, in our view, this type of bioinformatic annotation is inappropriate to interpretation of CpG sets obtained from machine-learning analysis. The elastic net algorithm we used to derive DunedinPACE is ignorant of biology. It simply chooses the minimum set of sites that, together, maximize prediction of the Pace of Aging criterion. For example, sites that are highly correlated statistically but which are located near different genes and therefore have potentially distinct biological interpretations are, from the view of the elastic net, interchangeable.

We note that the articles cited performed analysis on CpG sets compiled from many different measures of aging. In our view, this is a substantively different type of analysis than what we can conduct with the single set of probes included in DunedinPACE.

Pathway Analysis of DunedinPACE probes October 2021

We performed pathway analysis using the Ingenuity Pathway Analysis tool (IPA, Qiagen). We did this in three steps: First, we assigned DunedinPACE DNA methylation probes to the most proximal gene (in base pairs) based on genome assembly hg19. Second, we uploaded this list of genes to IPA, which mapped to 140 HUGO gene names (33 were unmapped and represented non-coding genes/ESTs). Of these, 7 gene names were duplicated (i.e. two DNA methylation probes mapped to the same gene); after removing the duplicates, 133 gene names were left for analysis. Third, we performed pathway analysis, restricting the search to (a) pathways known in humans, and (b) pathways experimentally observed in blood cells.

This analysis identified 9 canonical pathways which were enriched for genes in our list (nominal p-value cut-off = 0.05), with enrichment scores ranging from 1.29-1.86. These pathways were described by 12 of the 133 genes; Four genes were enriched in multiple pathways (GAD1, 4 pathways; ADCY8, 4 pathways; KCNQ2,2 pathways; SOCS3, 2 pathways).

Initial pathway analyses indicate that GABA synthesis and cell signaling might be pathways of interest with respect to DunedinPACE. However, there are some caveats to this conclusion. First, we selected the gene associated with a DNA methylation probe based on physical proximity; transcriptional relationships between genes and DNA methylation probes could be driven by factors other than physical distance between them. Such functional annotation could lead to gene assignation that differs from the one used here. Second, we do not have information on the relationship between the level of probe methylation and gene transcription; inclusion of this information could impact the assigned probabilities associated with pathway enrichment. The third caveat is that DunedinPACE is a statistical construct, not a biological one and, as such, inferences made about individual probes and their biological function may not be appropriate. It is important to remember that DunedinPACE, like PhenoAge and GrimAge, is an efficient algorithm that allows other researchers, who do not have the trained-upon data, to measure the phenotype of interest (in our case, the Pace of Aging which is made up of 20 years of repeatedly measured multiple biomarkers; in the case of GrimAge, a suite of plasma proteins, etc.). These algorithms comprise statistically efficient probes that capture a phenotype that would otherwise be inaccessible. Because their criteria for selection into the algorithm are based in statistics and not function, it is not appropriate to think of the resulting probes in biological terms. Moreover, DunedinPACE does not select every probe that is associated with the Pace of Aging. Rather, the elastic net has selected out probes that best predict DunedinPACE while minimizing both variance and bias in estimation. A biologically meaningful analysis could be performed through an Epigenome-Wide Association Study (EWAS) of the Pace of Aging. (The set of CpG sites with significant univariate associations with Pace of Aging is not equivalent to the set of elements identified by elastic net regression as most efficient for predicting Pace of Aging.) For these reasons, we are happy to share the results reported above in this response, but would prefer to not include them in the manuscript.

https://doi.org/10.7554/eLife.73420.sa2

Article and author information

Author details

  1. Daniel W Belsky

    Department of Epidemiology & Butler Columbia Aging Center, Columbia University, New York, United States
    Contribution
    Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review and editing
    For correspondence
    daniel.belsky@columbia.edu
    Competing interests
    is listed as an inventor on a Duke University and University of Otago invention that was licensed to a commercial entity
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-5463-2212
  2. Avshalom Caspi

    Center for Genomic and Computational Biology, Duke University, Durham, United States
    Contribution
    Conceptualization, Funding acquisition, Investigation, Project administration, Supervision, Writing – review and editing
    Competing interests
    is listed as an inventor on a Duke University and University of Otago invention that was licensed to a commercial entity
  3. David L Corcoran

    Center for Genomic and Computational Biology, Duke University, Durham, United States
    Contribution
    Formal analysis, Methodology, Software, Writing – review and editing
    Competing interests
    is listed as an inventor on a Duke University and University of Otago invention that was licensed to a commercial entity
  4. Karen Sugden

    Department of Psychology and Neuroscience, Duke University, Durham, United States
    Contribution
    Data curation, Methodology, Writing – review and editing
    Competing interests
    is listed as an inventor on a Duke University and University of Otago invention that was licensed to a commercial entity
  5. Richie Poulton

    Department of Psychology, University of Otago, Otago, New Zealand
    Contribution
    Data curation, Funding acquisition, Project administration, Resources, Writing – review and editing
    Competing interests
    is listed as an inventor on a Duke University and University of Otago invention that was licensed to a commercial entity
  6. Louise Arseneault

    Social, Genetic, and Developmental Psychiatry Centre, King's College London, London, United Kingdom
    Contribution
    Data curation, Writing – review and editing
    Competing interests
    No competing interests declared
  7. Andrea Baccarelli

    Department of Environmental Health Sciences, Columbia University, New York, United States
    Contribution
    Data curation, Writing – review and editing
    Competing interests
    No competing interests declared
  8. Kartik Chamarti

    Department of Psychology and Neuroscience, Duke University, Durham, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared
  9. Xu Gao

    Department of Occupational and Environmental Health, Peking University, Beijing, China
    Contribution
    Formal analysis, Writing – review and editing
    Competing interests
    No competing interests declared
  10. Eilis Hannon

    Complex Disease Epigenetics Group, University of Exeter, Exeter, United Kingdom
    Contribution
    Data curation, Writing – review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-6840-072X
  11. Hona Lee Harrington

    Department of Psychology and Neuroscience, Duke University, Durham, United States
    Contribution
    Data curation, Writing – review and editing
    Competing interests
    No competing interests declared
  12. Renate Houts

    Department of Psychology and Neuroscience, Duke University, Durham, United States
    Contribution
    Formal analysis, Writing – review and editing
    Competing interests
    No competing interests declared
  13. Meeraj Kothari

    Robert N Butler Columbia Aging Center, Columbia University, Brooklyn, United States
    Contribution
    Formal analysis, Visualization, Writing – review and editing
    Competing interests
    No competing interests declared
  14. Dayoon Kwon

    Robert N Butler Columbia Aging Center, Columbia University, New York, United States
    Contribution
    Formal analysis, Writing – review and editing
    Competing interests
    No competing interests declared
  15. Jonathan Mill

    Complex Disease Epigenetics Group, University of Exeter, Exeter, United Kingdom
    Contribution
    Data curation, Writing – review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-1115-3224
  16. Joel Schwartz

    Department of Environmental Health Sciences, Harvard TH Chan School of Public Health, Harvard University, Boston, United States
    Contribution
    Data curation, Writing – review and editing
    Competing interests
    No competing interests declared
  17. Pantel Vokonas

    Department of Medicine, VA Boston Healthcare System, Boston, United States
    Contribution
    Data curation, Writing – review and editing
    Competing interests
    No competing interests declared
  18. Cuicui Wang

    Department of Environmental Health Sciences, Harvard TH Chan School of Public Health, Harvard University, Boston, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared
  19. Benjamin S Williams

    Department of Psychology and Neuroscience, Duke University, Durham, United States
    Contribution
    Data curation, Writing – review and editing
    Competing interests
    No competing interests declared
  20. Terrie E Moffitt

    Department of Psychology and Neuroscience, Duke University, Durham, United States
    Contribution
    Conceptualization, Data curation, Funding acquisition, Project administration, Resources, Supervision, Writing – review and editing
    Competing interests
    is listed as an inventor on a Duke University and University of Otago invention that was licensed to a commercial entity

Funding

National Institute on Aging (AG032282)

  • Terrie E Moffitt

Medical Research Council (MR/P005918/1)

  • Terrie E Moffitt

National Institute on Aging (AG066887)

  • Daniel W Belsky

National Institute on Aging (AG061378)

  • Daniel W Belsky

New Zealand Health Research Council (16-604)

  • Richie Poulton

New Zealand Ministry of Business, Innovation and Employment (MBIE)

  • Richie Poulton

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

This research was supported by US-National Institute on Aging grants AG032282, AG061378, AG066887, and UK Medical Research Council grant MR/P005918/1.

The Dunedin Multidisciplinary Health and Development Research Unit is supported by the New Zealand Health Research Council Programme Grant (16-604), and the New Zealand Ministry of Business, Innovation and Employment (MBIE). We thank the Dunedin Study members, Unit research staff, and Study founder Phil Silva.

Understanding Society data come from The UK Household Longitudinal Study, which is led by the Institute for Social and Economic Research at the University of Essex and funded by the Economic and Social Research Council (ES/M008592/1). Information on how to access the data can be found on the Understanding Society website https://www.understandingsociety.ac.uk. Data governance was provided by the METADAC data access committee, funded by ESRC, Wellcome, and MRC (2015–2018: MR/N01104X/1; 2018–2020: ES/S008349/1).

The Normative Aging Study is supported by the National Institute of Environmental Health Sciences (grants P30ES009089, R01ES021733, R01ES025225, and R01ES027747). The VA Normative Aging Study is supported by the Cooperative Studies Program/Epidemiology Research and Information Center of the U.S. Department of Veterans Affairs competeand is a component of the Massachusetts Veterans Epidemiology Research and Information Center, Boston, Massachusetts.

The Framingham Heart Study is conducted and supported by the National Heart, Lung, and Blood Institute (NHLBI) in collaboration with Boston University (Contract No. N01-HC-25195, HHSN268201500001I and 75N92019D00031). This manuscript was not prepared in collaboration with investigators of the Framingham Heart Study and does not necessarily reflect the opinions or views of the Framingham Heart Study, Boston University, or NHLBI.

The E-Risk Study is supported by the UK Medical Research Council (grant G1002190), the US National Institute of Child Health and Development (grant HD077482), and the Jacobs Foundation. The generation of DNA methylation data was supported by the American Asthma Foundation.

This work used a high-performance computing facility partially supported by grant 2016-IDG-1013 (HARDAC+: Reproducible HPC for Next-generation Genomics") from the North Carolina Biotechnology Center. We would like to acknowledge the assistance of the Duke Molecular Physiology Institute Molecular Genomics Core for the generation of data for the manuscript.

AC, DC, DWB, KS, RP, and TEM are listed as inventors on a Duke University and University of Otago invention that was licensed to a commercial entity.

Senior Editor

  1. Jessica K Tyler, Weill Cornell Medicine, Cornell University, Ithaca, United States

Reviewing Editor

  1. Joris Deelen, Max Planck Institute for Biology of Ageing, Cologne, Germany

Reviewers

  1. Matthew Suderman, University of Bristol, United Kingdom
  2. Joris Deelen, Max Planck Institute for Biology of Ageing, Germany

Publication history

  1. Received: August 27, 2021
  2. Preprint posted: September 2, 2021 (view preprint)
  3. Accepted: December 13, 2021
  4. Accepted Manuscript published: January 14, 2022 (version 1)
  5. Version of Record published: February 17, 2022 (version 2)

Copyright

This is an open-access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication.

Metrics

  • 5,747
    Page views
  • 905
    Downloads
  • 22
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Daniel W Belsky
  2. Avshalom Caspi
  3. David L Corcoran
  4. Karen Sugden
  5. Richie Poulton
  6. Louise Arseneault
  7. Andrea Baccarelli
  8. Kartik Chamarti
  9. Xu Gao
  10. Eilis Hannon
  11. Hona Lee Harrington
  12. Renate Houts
  13. Meeraj Kothari
  14. Dayoon Kwon
  15. Jonathan Mill
  16. Joel Schwartz
  17. Pantel Vokonas
  18. Cuicui Wang
  19. Benjamin S Williams
  20. Terrie E Moffitt
(2022)
DunedinPACE, a DNA methylation biomarker of the pace of aging
eLife 11:e73420.
https://doi.org/10.7554/eLife.73420

Further reading

    1. Epidemiology and Global Health
    2. Medicine
    Qing Shen, Huan Song ... Unnur Valdimarsdóttir
    Research Article Updated

    Background:

    The association between cardiovascular disease (CVD) and selected psychiatric disorders has frequently been suggested while the potential role of familial factors and comorbidities in such association has rarely been investigated.

    Methods:

    We identified 869,056 patients newly diagnosed with CVD from 1987 to 2016 in Sweden with no history of psychiatric disorders, and 910,178 full siblings of these patients as well as 10 individually age- and sex-matched unrelated population controls (N = 8,690,560). Adjusting for multiple comorbid conditions, we used flexible parametric models and Cox models to estimate the association of CVD with risk of all subsequent psychiatric disorders, comparing rates of first incident psychiatric disorder among CVD patients with rates among unaffected full siblings and population controls.

    Results:

    The median age at diagnosis was 60 years for patients with CVD and 59.2% were male. During up to 30 years of follow-up, the crude incidence rates of psychiatric disorder were 7.1, 4.6, and 4.0 per 1000 person-years for patients with CVD, their siblings and population controls. In the sibling comparison, we observed an increased risk of psychiatric disorder during the first year after CVD diagnosis (hazard ratio [HR], 2.74; 95% confidence interval [CI], 2.62–2.87) and thereafter (1.45; 95% CI, 1.42–1.48). Increased risks were observed for all types of psychiatric disorders and among all diagnoses of CVD. We observed similar associations in the population comparison. CVD patients who developed a comorbid psychiatric disorder during the first year after diagnosis were at elevated risk of subsequent CVD death compared to patients without such comorbidity (HR, 1.55; 95% CI, 1.44–1.67).

    Conclusions:

    Patients diagnosed with CVD are at an elevated risk for subsequent psychiatric disorders independent of shared familial factors and comorbid conditions. Comorbid psychiatric disorders in patients with CVD are associated with higher risk of cardiovascular mortality suggesting that surveillance and treatment of psychiatric comorbidities should be considered as an integral part of clinical management of newly diagnosed CVD patients.

    Funding:

    This work was supported by the EU Horizon 2020 Research and Innovation Action Grant (CoMorMent, grant no. 847776 to UV, PFS, and FF), Grant of Excellence, Icelandic Research Fund (grant no. 163362-051 to UV), ERC Consolidator Grant (StressGene, grant no. 726413 to UV), Swedish Research Council (grant no. D0886501 to PFS), and US NIMH R01 MH123724 (to PFS).

    1. Epidemiology and Global Health
    Bingyi Yang, Bernardo García-Carreras ... Derek A Cummings
    Research Article

    Background: Over a life-course, human adaptive immunity to antigenically mutable pathogens exhibits competitive and facilitative interactions. We hypothesize that such interactions may lead to cyclic dynamics in immune responses over a lifetime.

    Methods: To investigate the cyclic behavior, we analyzed hemagglutination inhibition titers against 21 historical influenza A(H3N2) strains spanning 47 years from a cohort in Guangzhou, China and applied Fourier spectrum analysis. To investigate possible biological mechanisms, we simulated individual antibody profiles encompassing known feedbacks and interactions due to generally recognized immunological mechanisms.

    Results: We demonstrated a long-term periodicity (about 24 years) in individual antibody responses. The reported cycles were robust to analytic and sampling approaches. Simulations suggested that individual-level cross-reaction between antigenically similar strains likely explain the reported cycle. We showed that the reported cycles are predictable at both individual and birth-cohort level and that cohorts show a diversity of phases of these cycles. Phase of cycle was associated with the risk of seroconversion to circulating strains, after accounting for age and pre-existing titers of the circulating strains.

    Conclusions: Our findings reveal the existence of long-term periodicities in individual antibody responses to A(H3N2). We hypothesize that these cycles are driven by pre-existing antibody responses blunting responses to antigenically similar pathogens (by preventing infection and/or robust antibody responses upon infection), leading to reductions in antigen specific responses over time until individual's increasing risk leads to an infection with an antigenically distant enough virus to generate a robust immune response. These findings could help disentangle cohort-effects from individual-level exposure histories, improve our understanding of observed heterogeneous antibody responses to immunizations, and inform targeted vaccine strategy.

    Funding: This study was supported by grants from the NIH R56AG048075 (D.A.T.C., J.L.), NIH R01AI114703 (D.A.T.C., B.Y.), the Wellcome Trust 200861/Z/16/Z (S.R.) and 200187/Z/15/Z (S.R.). This work was also supported by research grants from Guangdong Government HZQB-KCZYZ-2021014 and 2019B121205009 (Y.G. and H.Z.). D.A.T.C., J.M.R. and S.R. acknowledge support from the National Institutes of Health Fogarty Institute (R01TW0008246). J.M.R. acknowledges support from the Medical Research Council (MR/S004793/1) and the Engineering and Physical Sciences Research Council (EP/N014499/1). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.