henhaobio
Sign in

bioRxivNeurosciencePreprint

Epigenetic and brain age across development: Performance and associations in the MIND consortium

Marlene StaginnusiD, Vilte BaltramonaityteiD, Isabel K. SchuurmansiD, Sarina AbrishamcariD, Martin Bauer, Sintia Iole BelangeroiD, Elisabeth B. BinderiD, Rodrigo A. BressaniD, Sybil Alexandra BurtiD, Claudia BußiD, Shi Yu ChaniD, Valentine ChirokoffiD, Shaunna Clark, H. Valerie CurraniD, Darina CzamaraiD, Serena DefinaiD, Kirsten Ann DonaldiD, Jules Roger DugreiD, Sonja EntringeriD, Johan G Eriksson, Janine F. FelixiD, Peter FransquetiD, Tom P. FreemaniD, Rodrigo Grassi‐OliveiraiD, Sorcha HamiltoniD, , Esther WaltoniD

PDFDOICodeData

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (30 Jul 2026) · CC BY 4.0

Across 15 cohorts (birth to 24 years), 20 epigenetic clocks and eight brain age models predicted chronological age modestly, stabilizing by middle childhood; their age residuals were weakly associated.

Epigenetic clocks and MRI brain age models have mostly been trained and applied in adults, leaving their validity in children and adolescents unclear. Earlier joint studies treated the two measures as separate outcome predictors, usually with a single brain age model and without youth samples; this work benchmarks 28 models together from birth to young adulthood and relates their residuals.

Key results

  1. ShownOn average, epigenetic clocks predicted chronological age only modestly in youth. Errors exceeded the sampled age range and varied widely between clocks. First-generation clocks erred least, especially the gestational clocks, which were applied only at birth. Fourth-generation clocks erred most.

    Pooled unweighted mean absolute error (MAE) 6.61 years (95% CI 2.78 to 15.70); pooled age-range-weighted MAE (MAEw) 2.18 (95% CI 0.74, 6.43); I^2 = 99.95%, 60.68% attributable to clock model; clock MAEw from 0.03 (Bohlin) to 27.34 (DamAge), QM(18) = 225166.02, pFDR < .001; pooled correlation with age r = .35 (95% CI .17, .50). · Performance of epigenetic clocks across…

  2. ShownMRI brain age models also had errors larger than the sample age range. Most of the variation came from cohorts, not from models. Pyment and DevBrainAge were the most accurate, and DunedinPACNI was the only model with a significant negative correlation with age.

    Unweighted MAE 4.46 years (95% CI 2.48, 6.43); pooled MAEw 1.72 (95% CI 0.67, 2.77); I^2 = 99.98%, with 94.40% from cohort and 5.42% from model; QM(7) = 7699.08, pFDR < .001; Pyment r = .34 (95% CI .25, .40); DunedinPACNI r = -.10 (95% CI -.19, -.02); overall r = .20 (95% CI .07, .32). · Performance of brain age models across…

  3. ShownEpigenetic clock accuracy improved with age and levelled off by middle childhood. Brain age models followed a similar trend, but it was not significant. A few clocks (PedBE, Horvath2013, skinHorvath) stayed comparatively stable across development.

    Epigenetic MAEw, gestational clocks excluded: b_log = -0.12, pFDR < .001, an 11.67% yearly MAEw decrease. Overall correlation with age did not change (b_z = 0.01, pFDR = .154). Brain age MAEw: b = -0.03 (95% CI -0.11, 0.05), pFDR = .484; correlation b_z = 0.01, pFDR = .242. · Developmental period; Developmental period

  4. ShownAs preregistered, models did better when the test cohort's age range overlapped the training age range. Epigenetic clocks also did better on the tissue they were trained on.

    Epigenetic geometric mean MAEw about 5.95 at 0% vs 3.04 at 100% age overlap (b_log = -0.01, pFDR < .001). Matching vs non-matching tissue: MAEw 1.92 vs 2.41 (b_log = -0.22, pFDR < .001), r .42 vs .31. Brain age MAEw about 2.01 at 0% vs 1.67 at 100% overlap (b = -0.003, pFDR < .001). · Age range overlap between training and testing…; Tissue match between training and testing data; Age range overlap between training and testing…

  5. ShownBrain and epigenetic predicted age residuals (PAR) were weakly and mostly positively associated, far less than residuals within each modality. Contrary to two of the hypotheses, the link did not depend significantly on age, clock generation (first vs next) or clock tissue.

    Pooled β = 0.017 (95% CI 0.01, 0.03), p = 0.001, over 136 model pairs (77.9% positive), vs unadjusted within-modality r = .25 (brain) and .22 (epigenetic). Moderators: age b = 0.001, pFDR = .477; generation b = 0.001, pFDR = .700; brain-tissue clocks b = 0.007, pFDR = .128. Largest pairs β = 0.06 (Kaufmann with skinHorvath; DevBrainAge with cAge). · Epigenetic PAR - brain PAR associations; Developmental period; Epigenetic clock generation; Tissue

Methods, data and reproducibility

Across 15 MIND cohorts (preregistered), 20 DNA methylation clocks and eight T1-weighted MRI brain age models were scored per cohort timepoint against chronological age (range-weighted MAE, Pearson r) and pooled by multilevel random-effects meta-analysis. Robust M-estimator regressions related brain to epigenetic age residuals, adjusting for age at MRI and methylation, sex and batch.

  • MIND (Methylation, Imaging, and NeuroDevelopment) consortium: international cohorts spanning six continents; 12 contributed repeated measures at up to six timepoints · human (cohort participants from birth to young adulthood) · n = 15 cohorts
  • DNA methylation from cord and whole blood, placenta, buccal and saliva samples (Illumina 450K and EPIC v1/v2 arrays), birth to 18 years · human · n = 20,917 observations across 33 timepoints
  • T1-weighted structural MRI (1.5T or 3T scanners), FreeSurfer-processed, ages 4 to 20 years · human · n = 19,925 observations across 30 timepoints
  • Observations with both DNA methylation and MRI, used for the residual association analyses · human · n = 5,559 observations across 19 timepoints
  • ALSPAC (Avon Longitudinal Study of Parents and Children), one contributing cohort; data available on request · human
Methods & tools
metafor (R package), FreeSurfer, Illumina Infinium HumanMethylation450K BeadChip, Infinium MethylationEPIC v1 and v2 BeadChip, Horvath2013, PedBE, DunedinPACE, cAge, Pyment, DevBrainAge, DunedinPACNI, Centile2
Software & environment
FreeSurfer version 6 or higher for MRI preprocessing; metafor R package for meta-analyses (version not stated in the main text).
Reproducibility gaps
Individual-level data are not public because of cohort consent and governance rules; access requires applying to each cohort (ALSPAC under proposal B3067). Preprocessing details and most per-model results are in the supplement, repository and dashboard.
Limitations (6)
  • Performance was judged only by error in predicting chronological age. Other benchmarks, such as association with age-relevant disease traits, may matter more for health questions and were not assessed; next-generation models might do better on them. · stated by the authors
  • Brain age could not be estimated in infants because T1 contrast is reversed and imaging features are hard to extract. In some analyses, early adulthood rested on a single dataset. · stated by the authors
  • Both epigenetic and brain age predictions were only moderately precise, so residuals derived from them may be unreliable. This could weaken the observed epigenetic-brain associations. · stated by the authors
  • Most clocks were trained largely on adult data, and no next-generation clock uses developmental training data. In development, both advanced and delayed aging may carry adverse effects that existing adult-trained clocks might not capture. · stated by the authors
  • Analyses were cross-sectional per timepoint, so age effects compare cohort-timepoint samples with different mean ages rather than tracking within-person change, even though 12 cohorts had repeated measures. · noted in this summary
  • Heterogeneity was extreme (I^2 = 99.95% for epigenetic and 99.98% for brain age performance), and 94.40% of the brain age heterogeneity came from cohorts. Pooled values therefore average over cohorts that differ in scanners, arrays, tissues and age ranges. · noted in this summary

Read this if you plan to apply epigenetic clocks or MRI brain age models in pediatric or adolescent cohorts and need evidence for choosing a model by age, tissue and training range.

Citations
1

Abstract

Understanding how biological age measures perform across development lays the groundwork for investigations into lifespan trajectories of healthy aging. We provide the most comprehensive assessment of epigenetic and brain age models across development (birth to 24 years; [≤]20,917 observations across 15 cohorts), evaluating how these models associate with chronological age and with each other, and how these associations change across development. Chronological age-prediction accuracy of epigenetic and brain age models was modest and varied substantially. Accuracy improved with age and stabilized by middle childhood. Few brain and fewer epigenetic clocks performed stably and well across all developmental stages. Performance was better when age range and tissue corresponded between training and testing data. Associations between epigenetic-brain age residuals were small, and changed little across development, tissues or clock generation. Given this developmentally dynamic system of epigenetic-brain age performances and associations, we give key recommendations to improve developmental research in this field.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISChan '26Ray '24Khoury '18Ball '21Shireby '20Haller '25
71% semantic matchNeuroscience
Shi Yu Chan, Pei Huang, Ai Ling Teh+16 more·bioRxiv·May 19, 2026