Review Papers Against Their Own Results

Visual syntheses — promises vs limitations

Figures below reproduce the paper's headline numerical summaries (correlations, absolute error proxies, classification accuracies) so you can visually apprehend where epigenetic clocks succeed and where they fail on ancient human methylomes (all numbers taken from Tawfik et al. 2026).

Data source: Tawfik et al. 2026 reporting median Monte Carlo Spearman correlations (interval-aware) per clock on corrected DamMet methylomes (10-CpG windows) Clock RMCS results (main paper)

Interpretation: only a small minority of clock-derived point estimates fall inside the anthropological intervals; Tawfik et al. report 2.8–13.9% (10‑CpG windows) depending on clock, indicating large absolute errors and systematic bias between damage-derived F-values and modern methylation scales Absolute accuracy limitations (main paper)

Note: Tawfik et al. show overall classification never exceeds ≈60% and Fisher tests indicate no significant association between predicted and anthropological age classes across windows — apparent class-imbalanced accuracies are driven by systematic window-dependent biases (overclassification to adults or subadults), not robust biological signal Classification results (main paper)

Critical evaluation (evidence-based, skeptical)

• Strengths
  • Careful dataset selection: high‑coverage ancient genomes (>10X) with USER-treated libraries and anthropological age labels; 36 individuals with depth ≈10–34X and sex labels — makes analyses focused and reproducible Sample selection (main paper)
  • Appropriate damage-aware pipeline: use of DamMet for probabilistic F-values, evaluation across multiple CpG window sizes (10–25 CpGs), and implementation of Liu‑style correction (RF + linear scaling) — consistent with field best-practices for damage-derived methylation inference DamMet method
  • Interval-aware evaluation: Monte Carlo sampling across anthropological age intervals for rank-based correlation rather than naive MAE, which respects uncertainty in osteological age labels and is statistically conservative Interval-aware evaluation (main paper)
• Limitations & potential biases (some noted by authors; some to emphasize):
  • Small ancient sample size and uneven age distribution: n=36 (and n=29 after filtering clear ages for classification) limits power for supervised retraining and cross‑validation; authors correctly emphasize instability and fold-sensitivity in mixed modern+ancient models Sample-size limitation (main paper)
  • Tissue bias — petrous bone dominance: petrous retains early-life matrix and low remodeling, possibly reflecting younger/homogenized methylation relative to chronological age at death; this undermines the assumption that methylation recovered from petrous reflects late-life epigenetic trajectories relevant to age-at-death Petrous bone tissue bias
  • Coverage limitations for CpG-level accuracy: DamMet and related methods require very high depth (>80X) for reliable per-CpG inference; ancient datasets (10–34X) force windowing/smoothing (10–25 CpGs) which reduces sensitivity to gradual age-related methylation changes and can introduce window-dependent biases (exactly what authors observe) DamMet depth requirements
  • Domain mismatch: modern clocks are trained on high-quality bisulfite/array/nanopore methylation data from living tissues; damage-derived F-values are a fundamentally different measurement domain (deamination-dependent inference, USER-treated biases, terminal damage effects). Correction helps but may not recover the exact measurement distribution or features informative for age models — authors show RF+linear rescaling improves Spearman R but not absolute age accuracy Correction improves similarity but not absolute accuracy
  • Potential confounders: post-mortem chemical alterations, library preparation heterogeneity, and differences in modern reference tissue (blood arrays) vs. ancient bone methylation remain sources of bias; authors partially control for these but full accounting requires bone-specific modern references and more varied ancient tissues Tissue/reference mismatch (main paper)

Where the evidence is strongest

1. Sex estimation from X‑chromosome methylation — robust and perfect in this dataset (PCA on ~4,000 X‑linked CpGs separates sexes) — consistent with literature that sex-specific methylation is a strong, global signal X-chromosome sexing success
2. Correction (Liu-style RF + linear scaling) reliably improves distributional similarity to modern references (Spearman R increases to ~0.6–0.8 depending on window), validating that damage-derived methylation contains recoverable structure — but structure is coarse and insufficient for precise age estimation Correction improves correlation but not full reconciliation

Key blindspots & what would falsify the paper's conclusions

The authors themselves outline falsifiable conditions: if a bone‑tissue‑specific ancient methylome reference and a model trained on genuinely ancient bone methylomes yielded robust continuous age predictions with narrow MAE across a substantial independent ancient test set, then the paper's core negative conclusion would be overturned. Practically, the missing pieces are (i) large, diverse ancient bone methylomes covering adult ages, (ii) higher per‑CpG coverage or better experimental assays (targeted capture of clock CpGs), and (iii) robust modern bone/bone‑age reference panels for correction. Tawfik et al. explicitly state this as the key path forward Authors

Concrete recommendations for next steps (experimental & computational)

• Targeted capture of published clock CpGs (e.g., Horvath/Universal clock CpGs) in USER‑treated ancient libraries to raise depth on informative sites — reduces noise without whole-genome >80X requirement.
• Create a modern bone methylation reference (petrous cortex and osteonal bone) across ages to enable domain-matched corrections and tissue-specific clocks; this is crucial because blood-derived references are imperfect proxies.
• Assemble a larger ancient bone methylome cohort (n≫36) with well-characterized osteological age ranges and multiple skeletal elements (teeth, long bones) to test tissue effects and to permit ancient-trained supervised clocks.
• Consider generative/domain-adaptation ML (e.g., adversarial domain adaptation) to learn mappings between damage‑derived F-values and modern beta values but only after collecting matched ancient-modern bone references to avoid overfitting and spurious corrections (see methodological cautions in epigenetic clock literature) Domain adaptation caveats in clock literature

Confidence & concluding judgement

Tawfik et al.'s negative result (current clocks cannot reliably estimate age-at-death from existing corrected ancient human methylomes) is well-supported by their analysis: they used appropriate pipelines (DamMet, correction, interval-aware metrics) and multiple modern clocks, and they report consistent failure modes across architectures and window sizes. The conclusion is robust at current data availability and technology; it does not rule out a future positive result when bone-specific ancient references, targeted assays, and larger sample sizes are obtained (i.e., the authors' recommendation is falsifiable and actionable). Full methodological details, code, and sample lists (Supplementary Table S1, methods) make the study reproducible in principle Reproducibility notes (main paper)