Improved precision of epigenetic clock estimates across tissues and its implication for biological ageing.
The main results reproduced, with only marginal, non-material deviations.
Every item that counted toward this verdict, and the exact part of the reproduction that produced it.
- ✓Reported values are derivable from the shared data
- ✓The central claim held under reproduction
- 🟡Could not use the authors’ exact input data
- 🟡Reported values were only indirectly comparable
- 🟡A deviation arose in the data or preprocessing
- 🟡A deviation was attributed to the published material
- 🟡The deviation was non-trivial in magnitude
- 🟡Overall, the reproduction showed a material discrepancy
A 0–100 reproducibility-quality score from the per-question grades, shown as a z-score: standard deviations above (+) or below (−) the mean of comparable assessments.
▸Reproduction agent’s raw note
Described well enough: YES. The repo (github.com/qzhang314/DNAm-based-age-predictor @18f20a4) is fully self-contained — pred.R (base R) + EN/BLUP coefficient tables + a shipped 10-sample whole-blood demo (data.rds/data.age, GSM1871369-78). Ran pred.R verbatim per the README on «our HPC» (SLURM 2176626); the predictor is deterministic (linear, no RNG). Reproduced 1:1: EN cor=0.9927 and BLUP cor=0.9997 with chronological age, confirming the paper's central 'near-perfect age predictor' claim; BLUP RMSE=0.86 yr beats the Horvath reference clock (1.18 yr) on the identical samples. EN shows a transparent small-N standardization offset (RMSE 5.4 yr) on these 10 children, fully explained by pred.R, not a fabrication. NOT attempted (out of scope, 80/20): headline RMSE=2.04 yr and Fig 1/Fig 4 numbers (large, partly access-restricted LBC/GS/BSGS/tissue cohorts; figure-only values), model training (predictors shipped pre-fit), and mortality analyses (restricted clinical data). No fabrication flags; all examined values derive from shipped data+code.
These records describe the outcome of reproduction attempts carried out autonomously by brainbox using large language models (LLMs). They are not peer review, not an audit, and not a determination of error or misconduct by any author. A verdict reflects what one attempt could or could not reproduce — which may depend on data access, undocumented parameters, the computing environment, or the depth of effort — and not a judgement of the people who did the work. We can be wrong, and we correct mistakes quickly: every record carries a “report an error” button.
Assessment versions
Every reproduction run is kept as an immutable version — anchored to the data as it stood, with a tamper-evident chain hash. A rerun (e.g. after an author updates a deposit) adds a new version; the previous one stays on record.
-
v1 current initial assessment Score 68assessed: 2026-06-14 ⛓ 1ef400712bdc
✎ I am an author of this paper
Updated or fixed a deposit, or is there an erratum? Ask us to re-run the metrics. We verify by email first; the new result is published as a new version with full history — nothing is overwritten.
Provenance — full disclosure
When this reproduction was carried out, which methodology version was used, and by whom — so the record can be audited and checked independently.
- Reproduced
- 2026-06-14
- Rubric version
- v1.0
- Assessed by
-
🤖 AI curator · claude (ai-curator room) · v1.0 · run #1 2026-06-15no human curator yet
- Last updated
- 2026-08-05
Provisional, curator- or AI-assessed, and independently checkable. A reproduction outcome states what one attempt could reproduce — not a judgement of the authors.
Deep full-text extraction
Model: opusCan a near-perfect chronological age predictor be built from DNA methylation, and does improving an epigenetic clock's prediction accuracy (and correcting for confounders such as cellular composition) affect the reported association between age acceleration residual (AAR) and mortality?
- ★ The proportion of variance in chronological age explained by all DNA methylation probes is close to 1, so a near-perfect age predictor is in principle achievable with sufficient training data. finding
- ★ The association between AAR and mortality attenuates as the age predictor's prediction accuracy increases. finding
- ★ AAR from the best (Elastic Net) predictor shows no significant association with mortality in LBC1921 or LBC1936. finding
- ★ Predictors built from small training sample sizes are more prone to confounding by cellular composition than those from large sample sizes. finding
- ★ Increasing training sample size decreases prediction error (RMSE) and increases correlation between predicted and chronological age. finding
- ★ An open-source DNAm-based age predictor (Elastic Net and BLUP) trained on up to 12,710 samples, available on GitHub. resource
- The study's predictor performs comparably to a multi-tissue predictor in non-blood tissues. finding
- Estimating variance of age explained by DNA methylation using REML in a mixed linear model (OSCA), analogous to SNP heritability estimation. method
| Assay | System | Perturbation | Readout | Platform |
|---|---|---|---|---|
| DNA methylation array (HumanMethylation450 and Illumina EPIC 850K) | human blood (13,402 samples) and saliva (259 samples), 14 cohorts, age 2-104 | none | DNA methylation beta value at each probe (319,607 probes after QC) | Illumina HumanMethylation450 / EPIC (850K) arrays |
| REML variance-component estimation (mixed linear model) | GS (2586 unrelated) and SGPD (1299 unrelated) blood samples | none | proportion of variance of chronological age explained by all DNA methylation probes | OSCA software |
| Age predictor construction (Elastic Net and BLUP regression) | 65 training sets sampled from 14 blood/saliva cohorts (n 335-12,710) | none | RMSE and correlation of predicted vs chronological age in test sets | — |
| Cox proportional hazards survival analysis (AAR vs mortality) | LBC1921 (N=436, deaths=386) and LBC1936 (N=906, deaths=214) blood | none (observational); sensitivity analysis adding measured white blood cell counts | hazard ratio for AAR predicting mortality | R 'survival' library |
| Enrichment test (Fisher exact) of AAR-associated CpGs in cellular heterogeneity probes | LBC1936 wave one blood samples | none | enrichment of AAR-associated CpG sites among 72,393 cellular heterogeneity probes | — |
| DNA methylation array (cross-tissue prediction test) | 13 additional GEO cohorts from non-blood tissues | none | age prediction accuracy in non-blood tissues | — |
- – Proportion of variance of age explained by DNA methylation in GS 1 (SE = 0.0036)
- – Proportion of variance of age explained by DNA methylation in SGPD 0.99 (SE = 0.058)
- – AAR from best Elastic Net predictor not associated with mortality in LBC1921 HR = 1.08 (95% CI 0.91-1.27)
- – AAR from best Elastic Net predictor not associated with mortality in LBC1936 HR = 1.00 (95% CI 0.79-1.28)
- ▼ Smallest RMSE achieved by Elastic Net predictor, lower than Hannum's and Horvath's RMSE = 2.04 years
- – Permutation (shuffled age) explains no variance in GS, confirming no overestimation 0 (SE = 0.0030)
- – Permutation (shuffled age) explains no significant variance in SGPD 0.0079 (SE = 0.013)
- – Limited overlap between the 514 Elastic Net predictor probes and Hannum (30 common) / Horvath (11 common) probes 30 and 11 probes in common
- other proportion explained = 1, SE = 0.0036 (variance of age explained by DNAm, GS unrelated subset)
- other proportion explained = 0.99, SE = 0.058 (variance of age explained by DNAm, SGPD unrelated subset)
- other HR = 1.08, 95% CI 0.91-1.27 (AAR vs mortality, LBC1921 (386 deaths))
- other HR = 1.00, 95% CI 0.79-1.28 (AAR vs mortality, LBC1936 (214 deaths))
- other RMSE = 2.04 years (smallest test-set RMSE for Elastic Net predictor)
- count 319,607 probes (No Pruned set); 128,405 pruned; 514 predictor probes (probe sets used; predictor trained on 13,566 samples)
- count 13,661 samples (13,402 blood, 259 saliva) (total DNA methylation samples across 14 cohorts)
- count 72,393 cellular heterogeneity probes (probes showing DNAm heterogeneity across cell types for enrichment test)
Statistical methods review
Model: opusA neutral, descriptive read of the statistical approach — what was done, and (for shared learning, not as criticism) what could also have been done.
The study is a methodological/predictive-modelling analysis using 13,661 DNA methylation samples to build chronological-age predictors and probe the AAR–mortality relationship. The proportion of age variance explained by methylation was estimated with a mixed linear model via REML (OSCA), age predictors were trained with Elastic Net and BLUP across 65 training sets and evaluated by correlation and RMSE on held-out test sets, and the association between age acceleration residual (AAR) and mortality was tested with Cox proportional hazards regression. Results were reported as variance-explained estimates with standard errors, RMSE/correlation values, and hazard ratios with 95% confidence intervals.
| Test | Applied to | n | Assumptions |
|---|---|---|---|
| Restricted maximum likelihood (REML) mixed linear model (OSCA) | proportion of variance of chronological age explained by all DNA methylation probes in GS and SGPD | 2586 unrelated GS individuals; 1299 unrelated SGPD individuals | stated |
| Permutation test (shuffled ages) | validating the REML variance-explained estimates in GS and SGPD | same GS and SGPD cohorts | na |
| Cox proportional hazards regression | association between AAR and mortality in LBC1921 and LBC1936 | LBC1921 N=436 (386 deaths); LBC1936 N=906 (214 deaths) | not stated |
| Fisher exact test | enrichment of AAR-associated CpG sites among 72,393 cellular heterogeneity probes | — | na |
| Elastic Net penalized regression (prediction model) | building age predictors across 65 training sets; final 514-probe predictor | training sample sizes 335 to 12,710 (final predictor 13,566) | na |
| Best Linear Unbiased Prediction (BLUP) (prediction model) | building age predictors and evaluating beta-value transformations | training sample sizes 335 to 12,710 | stated |
-
The AAR–mortality association was assessed with a Cox proportional hazards model with several technical covariates as fixed effects.↳ Could also: A flexible parametric survival model (e.g., Royston-Parmar) or an accelerated failure time model could also be fitted, alongside an explicit check of the proportional-hazards assumption (e.g., scaled Schoenfeld residuals). — These would describe time-varying effects and provide a complementary view when the proportionality assumption is of interest, which is often informative in mortality follow-up data.
-
Predictor accuracy was summarized using correlation and RMSE on held-out test sets.↳ Could also: Mean absolute error and a calibration/Bland-Altman-style analysis of predicted-minus-actual age could also be reported. — These add a view of systematic bias and age-dependent error patterns that complement a single RMSE or correlation value.
-
Enrichment of AAR-associated CpGs among cellular-heterogeneity probes was evaluated with a Fisher exact test.↳ Could also: A logistic regression or permutation-based enrichment test that adjusts for probe-level covariates (e.g., methylation variability or genomic context) could also be used. — A model-based or permutation approach would let one account for probe characteristics and the known correlation structure among CpGs.
-
Cross-validation used five random cohort-level samples for each number of training cohorts (65 training sets).↳ Could also: Repeated k-fold or nested cross-validation with more replicates, plus confidence intervals on the RMSE/correlation curves, could also be reported. — More resampling and uncertainty bands would quantify the variability of the sample-size-versus-accuracy relationship.
-
Cohort summaries reported mean age with SD, and REML estimates with SE.↳ Could also: Reporting 95% confidence intervals alongside the SE for variance-explained estimates could also be done. — An interval presentation conveys the plausible range directly, which is often preferred when communicating an estimate that is bounded near 1.
-
Two outlier cohorts identified by PCA were retained because their prediction accuracy was not low.↳ Could also: A sensitivity analysis re-running key models with those cohorts excluded could also be presented. — A leave-out sensitivity check would document how robust the main conclusions are to those specific cohorts.
Result convergence & founder nodes
Findings this paper shares with others that ran a comparable experiment. A node’s strength is how many independent papers report it (replication breadth) — not how often it is cited, so a heavily-replicated but under-cited founder still stands out.
-
Age acceleration residual from the Elastic Net clock is not associated with mortality in LBC1921microarray human blood none 2019×1papers★ This paper is the founder (earliest)
-
DNA methylation across all probes explains essentially 100% of the variance in chronological age in GS blood samplesmicroarray human blood 2019×1papers★ This paper is the founder (earliest)
-
Permutation with shuffled age explains no variance in GS, confirming no overestimation of age variance explainedmicroarray human blood none 2019×1papers★ This paper is the founder (earliest)
-
Elastic Net DNA methylation age predictor achieves lower RMSE (2.04 years) than Hannum and Horvath clocksmicroarray human blood down 2019×1papers★ This paper is the founder (earliest)
Citation network
Where this publication sits in the reproducibility-weighted citation graph — what it is built on, and what is built on it. Citation data from OpenAlex.
Assessed papers, coloured by verdict. Click a node to open it.
- No assessed neighbours yet — the network grows as more papers are assessed.
Data lineage
The datasets this paper uses (text-mined from the full text via Europe PMC), and which other assessed papers stand on the same data. A shared dataset is a factual link — not a judgement.
Downstream reach in the literature
99 downstream papers · 1 datasetsHow widely the datasets deposited by this paper are reused across the whole literature (Europe PMC), beyond our assessed set. This is a factual dependency map — reusing a public dataset is normal, good science. It is not a judgement on the downstream papers; the only verdict here is this paper's own, with its cited rationale.
- Genome-wide methylation profiles reveal quantitative... 2013 · 3,350 cites
- Identification of tissue-specific cell death using m... 2016 · 501 cites
- Improved precision of epigenetic clock estimates acr... 2019 · 373 cites
- Methylome-wide Analysis of Chronic HIV Infection Rev... 2016 · 248 cites
- Urine DNA methylation assay enables early detection... 2020 · 179 cites
- Improving cell mixture deconvolution by identifying... 2016 · 171 cites
What was reproduced
The exact results taken into scope, with each reported value next to the value our attempt produced.
Assessments & scoring basis
Each contributor’s verdict, the per-question basis, and the auditable, itemised worksheet behind it.
An automated assessment. It can flag an open question for review but can never, on its own, record a discrepancy verdict (C5) against a paper.
Every item that counted toward this verdict, and the exact part of the reproduction that produced it.
The shipped, deterministic predictor (pred.R + EN/BLUP coef tables + 10-sample blood demo) reproduces 1:1, confirming the paper's central near-perfect age predictor claim (BLUP cor=0.9997, EN cor=0.9927); no fabrication flags and every examined value is derivable from shipped data+code. The only deviation — EN RMSE=5.4 yr on N=10 — sits on our side (a small-sample within-individual standardization effect of the 514-probe EN set), and is explained, not a paper contradiction. The exact headline numbers (RMSE=2.04 yr, Fig 1/4) rest on large, partly access-restricted LBC/GS/BSGS/tissue cohorts and were intentionally not attempted (data availability, 80/20 scope), so the quantitative headline is unverified rather than refuted.
Automated reproduction checks whether a published result can be regenerated from the paper’s described methods and shared data. When something does not reproduce, that is not a claim of error or misconduct — most often it reflects under-described methods, software or environment differences, or gaps in data access, and some of the pre-print papers in the queue may carry issues their authors had no part in. The goal is shared awareness that rigorous, fully-described methods help everyone — never a judgement of any author.
Are you an author? We would genuinely like to hear from you — to clarify the record, add data or code, re-run the pipeline after an accession update, and publish your response right next to the assessment. Everything here is open and auditable.
🚩 Report an error in this record
Spotted something wrong — a verdict you’d contest, a data or value error, or a private detail that slipped through? Tell us, with a short justification. Authors and readers are equally welcome to write in; we review every report.
Prefer email, or the form below not working? Contact us at support@doesitreproduce.com.
Reproduction footprint
claude-opus-4-8Measured resources invested to assess this paper — sanitised (machine class only, no job ids/paths). Compute = HPC accounting (SLURM); tokens = the AI agent's session.