Improved epigenetic age prediction models by combining sex chromosome and autosomal markers.
The main result did not reproduce in this reproduction attempt. Where our recomputation produced values that differ from the published ones, those discrepancies are listed below. This is a single automated attempt — not peer review and not a finding of error or misconduct — and differences can also arise from data access, undocumented parameters or the computing environment. The verdict can be contested via “report an error”.
Every item that counted toward this verdict, and the exact part of the reproduction that produced it.
- ✓Reported values were directly comparable
- 🟡Could not use the authors’ exact input data
- 🔴A deviation arose in the data or preprocessing
- 🟡A deviation was attributed to the published material
- 🟡Reported values were not (fully) derivable from the shared data
- 🔴The deviation was non-trivial in magnitude
- 🔴The central claim did not (fully) hold under reproduction
- 🔴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 for a partial 1:1. Data (GEO GSE51032, EPIC-Italy buffy coat, 450K IDATs) is fully public; code (ZVWan/RandomForest_Age_Methylation) is unlicensed working R scripts with hardcoded paths, missing external probe-filter files, and no turnkey entrypoint, so we re-implemented the described method (minfi preprocessQuantile + detection-p + dropLociWithSnps; randomForest ntree=501 mtry=3; 2/3-1/3 split, 10 repeats) on GSE51032 alone on «our HPC» («job»). RESULT: the fully-specified autosomal-only baseline reproduced closely (RMSE 4.81 vs reported 5.06; MAD 3.89 vs 3.80 -> within-tol), validating data+method end-to-end on a larger 669-sample QC set (paper kept 336). The paper's CENTRAL claim that adding X-chromosome markers improves prediction did NOT reproduce with naive importance-based selection: X+autosomal gave RMSE 5.13, WORSE than autosomal-only (4.81) and worse than reported 4.51 -> mismatch. The improvement therefore depends on the repo's elaborate nested-CV / quadratic-derivative feature-count selection + curated 37-probe reduced set. NOT ATTEMPTED (hard 20%): the exact reduced-model 37-probe set / headline RMSE 2.54 y (flagged as implausibly low for blood 450K age prediction and a candidate for train/test leakage in the feature-selection loop -> auditor should check), the 4-dataset buffy-coat pooling, whole-blood (10-dataset) models, Y-only and sex-stratified sub-models.
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 48assessed: 2026-06-14 ⛓ 80e86d12793e
✎ 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: sonnetThe study hypothesized that incorporating sex chromosomal (X and Y) DNA methylation markers together with the most predictive autosomal DNAm markers would enhance epigenetic age prediction accuracy compared to autosomal markers alone.
- ★ Combining sex chromosomal DNAm markers with autosomal age-informative markers can produce a high-accuracy age prediction model competitive with autosomal-only models finding
- ★ Y chromosomal DNAm markers do not add predictive value beyond the six-autosomal-marker model finding
- ★ A reduced model combining 37 X chromosomal probes with 6 autosomal DNAm markers significantly improves age prediction accuracy (RMSE 2.54, MAD 1.89 years) finding
- ★ Four X chromosomal DNAm markers (cg27064949/DGAT2L6, cg04532200/PLXNB3, cg01882566/RPGR, cg25140188/intergenic) show strong correlation with chronological age finding
- Random forest regression is a robust, widely-used method for building DNAm-based age prediction models method
- ★ Sex-stratified and age-restricted data subsets yield model performance comparable to unstratified/unrestricted models finding
- A minimal six-marker autosomal DNAm panel from prior work serves as a reliable reference/positive-control model resource
| Assay | System | Perturbation | Readout | Platform |
|---|---|---|---|---|
| DNA methylation microarray (Infinium HumanMethylation450K) | human whole blood | none | DNAm beta values at CpG probes on sex chromosomes and autosomes | Illumina Infinium 450K BeadChip (GPL13534) |
| DNA methylation microarray (Infinium HumanMethylation450K) | human buffy coat | none | DNAm beta values at CpG probes on sex chromosomes and autosomes | Illumina Infinium 450K BeadChip (GPL13534) |
| Random forest regression age-prediction modeling with tenfold cross-validation | whole blood DNAm dataset | marker-set variation (X-only, Y-only, X/Y+autosomal, sex-stratified, age-restricted, reduced probe sets) | RMSE and MAD of predicted vs chronological age | randomForest v4.6.14 R package |
| Random forest regression age-prediction modeling with tenfold cross-validation | buffy coat DNAm dataset | marker-set variation (X-only, Y-only, X/Y+autosomal, sex-stratified, age-restricted, reduced probe sets) | RMSE and MAD of predicted vs chronological age | randomForest v4.6.14 R package |
| Spearman correlation analysis | whole blood and buffy coat DNAm data | none | correlation coefficient between DNAm level and chronological age | — |
- – Six-autosomal-marker model: RMSE 5.06y/MAD 3.80y (buffy coat) and RMSE 6.96y/MAD 5.24y (whole blood) RMSE 5.06-6.96 years
- ▼ Y-chromosome-only model performed poorly: RMSE 10.43y/MAD 8.41y (buffy coat), RMSE 14.29y/MAD 11.13y (whole blood) RMSE 10.43-14.29 years
- – Adding Y chromosomal markers to the 6 autosomal markers did not meaningfully improve performance over the autosomal-only model
- ▼ Reduced model with 37 X chromosomal + 6 autosomal markers achieved the best performance RMSE 2.54 years, MAD 1.89 years
- ▲ Four X chromosomal DNAm markers exhibited significant correlation with age Spearman r=0.50
- – Cross-validation across all sex-chromosome-based models yielded a range of error values RMSE 7.70-14.29y, MAD 6.10-11.13y
- – Sex-stratified and age-restricted models showed comparable RMSE/MAD to non-stratified, non-restricted models
- correlation Spearman r=0.50 (four X chromosomal DNAm markers vs chronological age)
- other RMSE=2.54 years, MAD=1.89 years (best reduced model: 37 X chromosomal + 6 autosomal markers)
- other RMSE=5.06y/MAD=3.80y (buffy coat) vs RMSE=6.96y/MAD=5.24y (whole blood) (six-autosomal-marker model performance)
- other RMSE=10.43y/MAD=8.41y (buffy coat) vs RMSE=14.29y/MAD=11.13y (whole blood) (Y-chromosome-only model performance)
- count 1291 whole blood samples, 547 buffy coat samples (total samples collected before QC/pre-processing)
- count 10,109 (buffy coat) and 10,178 (whole blood) X chromosome probes (probes retained after QC for X chromosome models)
- count 14 (buffy coat) and 35 (whole blood) Y chromosome probes (probes retained after QC for Y chromosome models)
- other RMSE 7.70-14.29 years, MAD 6.10-11.13 years (range across cross-validation of all constructed sex-chromosome-based models)
Statistical methods review
Model: sonnetA 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 built random forest regression (RFR) models from publicly available Illumina Infinium 450K DNA methylation data to predict chronological age using sex chromosomal and autosomal CpG markers in whole blood and buffy coat. All models were evaluated via 10-fold cross-validation, with performance reported as RMSE ± 95% CI and MAD ± 95% CI. Spearman rank correlations between individual CpG beta-values and chronological age were computed to validate the top-ranked markers from the RFR feature-importance output. Comparisons between model types (e.g., 'significantly improved accuracy') were made by inspecting point estimates and 95% CI ranges rather than by formal inferential tests.
| Test | Applied to | n | Assumptions |
|---|---|---|---|
| Random forest regression (RFR) with 10-fold cross-validation; hyperparameters ntree=501, mtry=3 | All age-prediction models: Y-only, X-only, combined sex chromosome + autosomal, sex- and age-stratified subsets, and 50 reduced models (1–50 probes) | 860 training / 481 test (whole blood); 365 training / 182 test (buffy coat); 2:1 split | not stated |
| Spearman rank correlation (DNAm beta-value vs. chronological age) | Post-hoc validation of top 15 age-informative CpG markers per model; threshold rs ≥ 0.50 cited for four X chromosomal markers | — | not stated |
| Detection p-value test (minfi package) | Probe-level quality control; probes with p-detection > 0.01 excluded | — | not stated |
| Statistical significance test (p < 0.05) for cell-type differences | Probe filtering step to exclude probes showing significant methylation differences between whole blood and buffy coat cell compositions | — | not stated |
-
Random forest regression was the sole prediction algorithm evaluated↳ Could also: Regularized linear regression (elastic net or LASSO, as used in the original Horvath and Hannum clocks) or gradient boosting (e.g., XGBoost) could also be applied — Elastic net / LASSO produce explicit linear coefficients, making the contribution of each CpG directly interpretable and facilitating comparison with published clock weights; gradient boosting often achieves competitive accuracy and can provide calibrated prediction intervals
-
Model-to-model accuracy comparisons (e.g., 'significantly improved accuracy') were based on inspecting RMSE and MAD point estimates without a formal statistical test↳ Could also: A paired permutation test or Diebold-Mariano test on fold-level prediction errors from the same cross-validation folds could also formally compare models — Formally testing prediction-error differences provides an explicit p-value and controls the chance of treating a difference as meaningful when it falls within the sampling variability of cross-validation folds
-
Spearman correlations were computed for individual top-ranked CpGs without multiple testing correction↳ Could also: Benjamini-Hochberg FDR correction (or Bonferroni) could also be applied across all CpG-age associations tested — When correlations are evaluated for many probes simultaneously, FDR or family-wise error rate correction calibrates the expected proportion of false discoveries; this is standard practice in array-wide association analyses
-
A nominal p < 0.05 threshold was applied across tens of thousands of probes in the cell-type difference filtering step↳ Could also: A Benjamini-Hochberg FDR-adjusted threshold (q < 0.05) could also be applied at this filtering step — With tens of thousands of simultaneous tests, a nominal p < 0.05 threshold yields a large expected number of false exclusions; FDR adjustment would calibrate the filtering to a specified false discovery rate
-
Performance was summarized with RMSE and MAD only↳ Could also: Pearson r between predicted and chronological age, and age-acceleration bias plots (residual vs. chronological age), could also be reported — Pearson r is widely reported in epigenetic clock literature and aids cross-study comparability; bias plots quantify the systematic over- or under-prediction at young and old age extremes that the authors note qualitatively but do not report numerically
-
Hyperparameters (ntree, mtry) were tuned by grid search within the same cross-validation loop used for performance estimation↳ Could also: Nested cross-validation (inner loop for tuning, outer loop for evaluation) could also be used — Nested CV separates hyperparameter selection from generalization estimation, which avoids the potential for optimistic bias when the same data inform both the tuning and the reported error
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.
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.
What was reproduced
The exact results taken into scope, with each reported value next to the value our attempt produced.
Scope — pmid-40665390
Paper: Wan Z et al. (2025) Improved epigenetic age prediction models by combining sex chromosome and autosomal markers. Epigenetics & Chromatin. DOI 10.1186/s13072-025-00606-5 · PMCID PMC12261677.
Code: https://github.com/ZVWan/RandomForest_Age_Methylation
(commit 12066659561e35207c14c2c552616a830377f2db, 2022-09-29; 5 R scripts, no
license, no turnkey entrypoint — working scripts with hardcoded paths/objects).
Data assigned to this RU: GEO GSE51032 = "EPIC-Italy at HuGeF", 845 buffy-coat samples, Illumina 450K (GPL13534), age + gender in sample metadata, IDAT-only (no processed beta matrix shipped → minfi preprocessing required). In the paper GSE51032 is the dominant buffy-coat dataset (336 of 547 post-QC buffy-coat samples; pooled with GSE55491/GSE61496/GSE74548).
Method as described (Methods + repo)
- minfi QC + normalisation of 450K IDATs (preprocessQuantile/Funnorm), sex prediction, drop bad samples.
- Probe filtering: keep X/Y-chromosome probes + 6 named autosomal age markers; drop SNP-overlap (dropLociWithSnps), Chen-2013 cross-hybridising probes, Jaffe blood-cell-type probes, detection-p failures.
- Random forest age regressor: ntree=501, mtry=3, tenfold CV, 2/3 train / 1/3 test, feature ranking by RF importance, keep top ~15 probes (reduced model).
- Report RMSE and MAD of predicted vs chronological age per model.
6 autosomal markers (whole-blood set, from repo Probe_Quality_..._Preprocessing.R):
cg16867657, cg12934382, cg11807280, cg02872426, cg06874016, cg08097417
(ELOVL2, etc. — the Zbieć-Piekarska 2017 forensic age CpGs).
IN SCOPE (pipeline-derived; we attempt)
- C1 — Autosomal-only baseline (buffy coat). RF age model on GSE51032 using the 6 autosomal markers only. Reported buffy-coat baseline: RMSE 5.06 y, MAD 3.80 y (Results / Table S1). Primary clean 1:1 data point — fully specified, tiny, low-ambiguity.
- C2 — X-chromosome + autosomal (buffy coat). RF age model adding all filtered X-chromosome probes to the 6 autosomal markers, importance-selected. Reported full X+autosomal buffy-coat: RMSE 4.51 y, MAD 3.42 y (Table S1). Tests the paper's central claim direction (does adding X markers beat autosomal-only?). Secondary, more ambiguous.
OUT OF SCOPE / NOT chased (the hard ~20%)
- Exact reduced-model headline (37 X + 6 autosomal → RMSE 2.54 y / MAD 1.89 y): depends on the exact 37-probe set selected by the repo's quirky quadratic-derivative feature-count selection over a 4-dataset pool — not reproducible 1:1 from GSE51032 alone with reasonable effort.
- Whole-blood models (10 datasets, 1291 samples) — different tissue pool.
- Pooling all 4 buffy-coat datasets — we use GSE51032 alone (336/547); our sample set therefore differs from theirs (and their 845→336 QC reduction is itself under-documented). Comparison is honest-but-not-identical-cohort.
- Y-chromosome-only / sex-stratified sub-models, all figures/plots.
Why this is a valid reproduction
Per brief P16, running the described RF method on the paper's own data per the stated parameters is equally valid. C1 is a clean, fully-specified data point; C2 probes the headline claim's direction. We do not present grades as ground truth — a human auditor decides.
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 fully-specified autosomal baseline (C1) reproduces cleanly (RMSE 4.81 vs 5.06 y; MAD 3.89 vs 3.80 y), validating the public GSE51032 data and basic RF method end-to-end. However, the paper's central claim — that adding sex-chromosome markers improves prediction — did not reproduce: our X+autosomal model (5.13 y) was worse than autosomal-only (4.81 y), reversing the reported gain (5.06→4.51 y). This is partly our doing (we used naive importance selection, not their nested-CV pipeline), but the flagship RMSE 2.54 y is implausibly low for blood 450K age prediction and is consistent with train/test leakage in the feature-selection loop. Net: baseline solid, but the headline improvement is unconfirmed and leakage-suspect — a substantive, documentation-worthy discrepancy on the authors' core conclusion.
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.