Bio Methylation Epigenetic Clocks

Version Compatibility

Reference examples tested with: methylclock 1.8+, dnaMethyAge (GitHub yiluyucheng), methylCIPHER (GitHub MorganLevineLab), DunedinPACE (GitHub danbelsky).

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('<pkg>') then ?function_name to verify parameters

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

The clock coefficient sets are FIXED and version-pinned (a clock is a frozen list of CpGs and weights), so the package version mostly controls which clocks ship and what the clock-name strings are. Verify accepted names live: methylclock via checkClocks(beta); dnaMethyAge via availableClock(). The ARRAY PLATFORM is the version that matters most: EPICv2 drops a clock-specific fraction of CpGs, so always report how many of each clock's CpGs were actually present.

Epigenetic Clocks

"How old is this sample epigenetically?" -> Apply a frozen elastic-net clock to the beta matrix, then report the age ACCELERATION (residual vs chronological age), not the raw age - because a clock is a predictor, and the residual is the signal.

• R: DNAmAge(beta, clocks = c('Horvath', 'Hannum', 'Levine'), age = pheno$age)

Scope: applying pre-trained clocks (DNAm age, pace, mitotic) and computing age acceleration from a clean beta/M-value matrix. Cell-count adjustment for IEAA -> cell-type-deconvolution. Clean beta matrix / EPICv2 replicate-probe collapse -> array preprocessing. Training a predictor / cross-validation / leakage -> machine-learning/model-validation. Survival/mortality modeling of EAA -> clinical-biostatistics/survival-analysis. Per-CpG and region testing -> differential-cpg-testing, dmr-detection.

The Single Most Important Modern Insight -- A Clock Is a Predictor, Not a Mechanism, and Not Even a Reliable One Per-CpG

DNAm age is a frozen elastic-net weighted sum over CpGs that a penalty chose for out-of-sample prediction. The CpGs are prediction features, never an aging pathway. Four corollaries every common misuse violates:

1. The CpG set is not biology. Do not GO-enrich clock CpGs. Two clocks for the same outcome can share almost zero CpGs (the elastic net arbitrarily keeps one of many correlated predictors), so non-overlap is expected, not a contradiction.
2. The endpoint is age ACCELERATION, not the raw age. Raw DNAm age just recapitulates chronological age (r often > 0.9). The signal is the residual of DNAm age on chronological age (EAA); IEAA additionally residualizes on cell counts - which is exactly where deconvolution meets this skill.
3. First-gen per-CpG reliability can be smaller than the effect being chased. Many first-gen clock CpGs have low test-retest ICC (Sugden 2020 Patterns 1:100014), so the same sample can age several years between technical replicates. PC clocks (Higgins-Chen 2022 Nat Aging 2:644) exist specifically to fix this for longitudinal and trial use.
4. Association is not causation is not transfer. EAA associating with an exposure does not make the clock causal; a blood clock does not automatically work in another tissue or ancestry without a recalibration check.

Organize the analysis around defending these four, not around listing clock names.

The Clock Menu by Question

The question dictates the clock; there is no single best clock. Pick by what is being predicted, not by popularity.

DunedinPACE is reported as the raw PACE value (already a rate); never residualize it like an age clock and never compare its number to Horvath years.

Decision Tree by Scenario

Apply a Clock and Extract Age Acceleration

Goal: Compute DNAm age for several clocks and turn the raw ages into age acceleration, the actual endpoint.

Approach: Run DNAmAge with chronological age supplied so it returns acceleration columns directly; ageAcc is the raw DNAm-minus-chronological difference and ageAcc2 is the residual of DNAm age on chronological age (the EAA to test). Always check clock-CpG coverage first.

library(methylclock)

cpg_report <- checkClocks(beta)        # which clock CpGs are missing BEFORE estimating
ages <- DNAmAge(beta, clocks = c('Horvath', 'Hannum', 'Levine', 'skinHorvath'),
                age = pheno$age,       # supplying age yields ageAcc and ageAcc2 columns
                cell.count = FALSE,    # set TRUE only when adjusting toward IEAA-style estimates
                min.perc = 0.8)        # refuse a clock missing >20% of its CpGs (default 0.8)
# ageAcc  = DNAm age - chronological age (raw difference)
# ageAcc2 = residual of DNAm age on chronological age = the EAA endpoint with cell.count=FALSE
#   (methylclock labels ageAcc2 "similar to IEAA"; confirm the exact column semantics in the
#    installed vignette, and the cell-count-adjusted residual when cell.count=TRUE)

The dnaMethyAge package returns acceleration in one call and exposes author-year clock IDs:

library(dnaMethyAge)
availableClock()                                  # confirm the installed clock-name strings
phenoage <- methyAge(beta, clock = 'LevineM2018',
                     age_info = pheno,            # data.frame with Sample, Age (Sex for GrimAge variants)
                     fit_method = 'Linear')       # adds an Age_Acceleration column

Pace of Aging Is Not an Age

Goal: Compute DunedinPACE as a rate and keep it on its own scale.

Approach: Use the dedicated package; the output is a per-sample pace (~1.0 = normal). Do not regress it on chronological age and do not merge it with age-clock acceleration.

library(DunedinPACE)
pace <- PACEProjector(beta)   # returns the DunedinPACE pace values (~1.0 = normal, >1 = faster aging); report as-is
# Never residualize PACE on chronological age and never compare its value to Horvath years.

Reliability and PC Clocks

For longitudinal, interventional, or clinical-trial endpoints, first-gen per-CpG noise (Sugden 2020) can swamp a small intervention effect. PC clocks (Higgins-Chen 2022) train the elastic net on principal components across thousands of CpGs, averaging out per-CpG noise and lifting test-retest ICC toward ~0.9. They live in methylCIPHER (MorganLevineLab), not in methylclock. Use a PC clock, or at minimum document an ICC/reliability assessment, for any repeated-measures design.

Per-Method Failure Modes

GO-enriching clock CpGs

Trigger: running pathway enrichment on a clock's CpG list to "explain aging." Mechanism: the CpGs are penalty-selected prediction features, one arbitrary representative per correlated cluster. Symptom: a plausible-looking enrichment that is an artifact of feature selection. Fix: do not enrich clock CpGs; treat them as predictors only.

Reporting raw DNAm age instead of acceleration

Trigger: correlating raw DNAm age with an exposure. Mechanism: raw age is dominated by chronological age (r > 0.9). Symptom: every clock "associates" with age-correlated variables. Fix: test the residual (ageAcc2 / Age_Acceleration / IEAA), not the raw age.