Table 4.

Description of DNA-methylation algorithm computations

DNAm algorithm	Description
PhenoAge	PhenoAge was first modeled from physiological markers and chronological age [28]. This first-stage algorithm was then applied to a new sample in which it was modeled from DNA methylation to derive the final DNA-methylation clock. PhenoAge represents the age in years at which average mortality risk in NHANES III matches the mortality risk predicted by the PhenoAge algorithm PhenoAge was computed using DNAm principal components, which have been found to increase reliability [30], using code available at https://github.com/MorganLevineLab/PC-Clocks. Using 24 technical replicates of samples in SOEP, we estimated the intraclass correlation coefficient (ICC). PhenoAge showed excellent reliability (ICC = 0.982). PhenoAge Acceleration was computed by residualizing PhenoAge for chronological age
GrimAge	GrimAge was developed with a set of physiological indicators modeled from DNAm using machine learning analysis, and then these DNA-methylation algorithms along with age, sex, and a DNAm algorithm of smoking history were applied to model mortality [29]. GrimAge represents the age in years at which average mortality risk in the Framingham Heart Study Offspring cohort matches predicted mortality risk GrimAge was computed using DNAm principal components, which have been found to increase reliability [30], using code available at https://github.com/MorganLevineLab/PC-Clocks. GrimAge showed excellent reliability (ICC = 0.999). GrimAge Acceleration was computed by residualizing GrimAge for chronological age
DunedinPACE	DunedinPACE was developed as a DNA-methylation measure of the pace of aging in the Dunedin Study birth cohort [17]. The Dunedin Study Pace of Aging is a composite phenotype derived from analysis of longitudinal change in biomarkers of organ-system integrity. Initially developed from analysis of three waves of biomarker data accumulated over a 12-year period [31]. Pace of Aging has recently been extended to a fourth measurement occasion spanning 20 years of follow-up [32]. DunedinPACE was developed from this second iteration of the Pace of Aging Briefly, DNAm algorithm development was conducted using a subset of EPIC array probes that were also included on Illumina’s earlier 450 k array and that were identified as having relatively higher test–retest reliability [33]. Elastic net regression machine learning analysis was used to fit Pace of Aging to DNAm data generated from blood samples collected when participants were aged 45 years. The elastic net regression produced a 173-CpG algorithm. Increments in DunedinPACE correspond to “years” of physiological change occurring per 12-months of chronological time. A value of 1 reflects the average Pace of Aging in the Dunedin Study birth cohort over the age 26–45 follow-up period. A value of 1.01 therefore reflects a pace of aging 1% faster than the Dunedin Study norm. DunedinPACE was be calculated based on the published algorithm using code available at https://github.com/danbelsky/DunedinPACE/. Fourteen of the 173 CpG probes that are part of DunedinPACE were not present in our dataset. Buccal DunedinPACE showed good reliability (ICC = 0.74)
Epigenetic-g	Epigenetic-g was computed using a blood-based algorithm from an epigenome-wide association study (EWAS) in BayesR + of general cognitive function (g) in 9162 adults (59% females; mean age 49.8 years, SD 13.6, range 18–93) in the Generation Scotland Study [11]. Briefly, general cognitive function was derived from the first unrotated principal component of logical memory, verbal fluency and digit symbol tests, and vocabulary. Cognitive phenotypes were corrected for age, sex, BMI and an epigenetic smoking score. Epigenetic-g includes all CpG sites in the EWAS. The weights for each CpG are the mean posterior effect sizes from the EWAS model of g. Prior to computation of Epigenetic-g in the present study, methylation values were scaled within each CpG site (mean = 0, SD = 1) and calculated based on the published algorithm using code available at https://gitlab.com/danielmccartney/ewas_of_cognitive_funct. Epigenetic-g showed good reliability (ICC = 0.84)
PedBE	As a data quality control, we examined associations of chronological age with the Pediatric-Buccal-Epigenetic (PedBE) clock, which was developed to predict chronological age in individuals aged < 20 years from buccal-cell DNAm, i.e., the same tissue type examined here [34]. While the pediatric sample used to develop the PedBE clock is considerably younger than our lifespan sample, it is one of the few aging-related DNAm indicators developed using buccal cells. Correspondence between PedBE and chronological age in our sample increases confidence in the quality of our buccal-cell DNAm data Elastic net penalized regression was used to select 94 CpGs from a training dataset of 1032 subjects. PedBE was calculated based on the published algorithm using code available at https://github.com/kobor-lab/Public-Scripts/blob/master/PedBE.Md. All 94 CpG probes were present in our dataset. PedBE showed excellent reliability (ICC = 0.967). PedBE was strongly associated with chronological age, indicating good data quality (r = 0.91, 95% CI = 0.90, 0.92, p < 0.001)