Reduced DNA methylation patterning and transcriptional connectivity define human skin aging

Reduced DNA methylation patterning and transcriptional connectivity define human skin aging (Bormann et al. 2016)

Bormann F, Rodríguez-Paredes M, Hagemann S, et al. · Aging Cell 15(3):563–571 · 2016 · DOI: 10.1111/acel.12470

13 authors. PMID: 27004597. PMC: PMC4854925. Gold open access. Citation count: ~94 (OpenAlex, 2026-05-19; 100th percentile FWCI 4.85). Affiliated institutions include German Cancer Research Center (DKFZ), Beiersdorf AG, and University of Greifswald.

TL;DR

Bormann et al. 2016 is the founding study of epidermis-specific DNA methylation clocks. Using 108 female Caucasian epidermis samples (combined punch-biopsy and suction-blister cohorts), the authors trained a support vector machine (SVM) predictor on the Illumina 450k array that estimates chronological age with MAE <5.25 yr — outperforming the pan-tissue Horvath 2013 clock (MAE 12.1 yr on the same independent test set, a 2.3-fold improvement). Beyond clock construction, the paper characterizes a mechanistic epigenetic-aging signature in the epidermis: age reduces the dynamic range of methylation (CpGs lose sharp fully-methylated vs. fully-unmethylated boundaries), increases inter-individual methylation heterogeneity, and reduces transcriptional network connectivity in old versus young epidermis. These three features together are interpreted as evidence for loss of epigenetic regulatory fidelity as a defining feature of aging skin. The training cohort of 108 white females and the 450k probe set were later re-used by Qi 2026 to build the 23k/173-CpG tissue-specific clock validated cross-ethnically.

Background

Why tissue-specific clocks matter

The Horvath 2013 pan-tissue clock (353 CpGs; MAE 3.6 yr held-out across 51 tissues) ¹ was a major advance, but its cross-tissue portability came at the cost of within-tissue accuracy. When applied to tissue-specific samples that were not well-represented in its training data, or to non-blood tissues with distinct cell-type compositions, the pan-tissue clock underperforms. Epidermis presents a particularly sharp case: it is a stratified squamous epithelium (predominantly keratinocytes) with a completely different epigenetic landscape from the hematopoietic and stromal cells that dominate Horvath’s training set. Bormann et al. asked whether an epidermis-dedicated model could substantially outperform the pan-tissue predictor.

Epidermis also presents practical advantages for biological-age measurement: it is non-invasively accessible (tape-stripping, suction blister) and undergoes continuous self-renewal from epidermal stem cells — meaning its methylation state reflects ongoing epigenetic drift rather than accumulated post-mitotic damage. Any tissue-specific clock trained here is simultaneously a tool for measuring skin-specific aging and a potential readout for interventions targeting epidermal biology.

Methods

Cohort design (two independent datasets, combined n=108)

Dataset	n	Age range	Sex	Ethnicity	Sampling method	Anatomical site
Dataset 1 (discovery)	48	18–27 yr (young, n=24); 61–78 yr (old, n=24)	Female	Caucasian	Punch biopsy	Outer forearm
Dataset 2 (extended)	60	20–79 yr (continuous)	Female	Caucasian	Suction blister roof	Outer forearm
Combined training	108	18–78 yr	Female	Caucasian	Both	Outer forearm

Note: both datasets are sex- and ethnicity-restricted to female Caucasian donors. This is a key generalizability limitation (see § Limitations).

For clock training, the young/old dichotomy in Dataset 1 was used for differential methylation analysis; the continuous age range in Dataset 2 and combined n=108 were used for regression modeling and leave-one-out / 10-fold cross-validation.

Array platform and preprocessing

Illumina HumanMethylation450 BeadChip (450k array; >450,000 CpG sites measured). Epidermis DNA was extracted from the biopsy or blister specimens; standard Illumina preprocessing and quality filtering applied (minfi package; sex-chromosome probes removed; cross-reactive probes and those with allele frequency ≥0.01 within 5 bp of the SBE site removed). The final SVM was trained using the complete set of 450k probes as independent variables — the paper explicitly states this (Methods, p.565); no post-filter probe-count ambiguity applies.

Prediction model

Support vector machine (SVM) regression trained against chronological age on the combined n=108 dataset. Model selection: leave-one-out cross-validation (LOOCV) and 10-fold cross-validation. The authors do not describe a final held-out test set for the SVM training cohort itself — the primary performance metrics come from applying the trained model to the independent Vandiver et al. 2015 dataset (n=18 sun-exposed forearm epidermis samples), which was completely separate from training.

This is an important distinction: the MAE <5.25 yr figure is from the leave-one-out cross-validation within the training set; the Vandiver validation (MAE 6.72 yr) is the true out-of-sample performance.

Transcriptome and methylation-pattern analyses

In addition to clock modeling, the paper performed:

• Intramethylome variance analysis: compared the within-sample variance of methylation beta values between young and old groups (P = 0.0016 for reduced variance in old samples).
• Spatial CpG correlation: measured the spatial autocorrelation of methylation among neighboring CpG sites (P = 5.4×10⁻⁷ for reduced spatial correlation with age).
• Intermethylome heterogeneity: compared the pairwise similarity of methylomes between young individuals versus between old individuals.
• Transcriptional network connectivity: RNA-seq of paired epidermis samples; built co-expression networks; quantified the fraction of gene pairs with significant expression correlation in young versus old.

Results

1. Clock performance — training and cross-validation

Metric	Value	Method
Pearson ρ (training, LOOCV)	0.92 (p < 2.2×10⁻¹⁶)	Leave-one-out cross-validation, n=108
MAE (training, LOOCV)	<5.25 yr	Leave-one-out cross-validation
10-fold cross-validation	Concordant with LOOCV	See paper supplementary

2. Independent validation (Vandiver 2015 cohort, n=18 sun-exposed)

Metric	Bormann SVM	Horvath pan-tissue clock
Pearson ρ	0.96 (p = 1.66×10⁻¹⁰)	0.89
R²	0.93	0.794
MAE	6.72 yr	12.1 yr

The Bormann epidermis clock reduces MAE from 12.1 yr to 6.72 yr on independent sun-exposed epidermis — a 1.8-fold improvement over Horvath in this tissue. The Vandiver cohort (Vandiver et al. 2015, PMC4506614) consisted of 18 sun-exposed forearm epidermis samples with known donor ages; it was not used in training.

3. Methylation pattern erosion with age

Aging epidermis shows a consistent shift in the architecture of DNA methylation, beyond simple gain or loss at individual CpGs:

• Reduced dynamic range: young methylomes show sharply demarcated regions of near-complete methylation (β ≈ 1) versus near-zero methylation (β ≈ 0). Aged methylomes show intermediate β values more frequently — the bimodal distribution characteristic of young methylomes flattens toward a unimodal intermediate distribution. This erosion of binary methylation states is quantified as reduced intramethylome variance (P = 0.0016).
• Reduced spatial autocorrelation: neighboring CpGs in young methylomes are strongly correlated (either both high or both low β). In aged methylomes, this spatial correlation decays — neighboring CpGs become more independent in their methylation state (P = 5.4×10⁻⁷ for reduced spatial correlation with age). This reflects physical disruption of coordinated methylation domains rather than random stochastic drift, because the decay is spatially structured.
• Increased intermethylome heterogeneity: young individuals have highly similar methylomes to each other (tight clustering); old individuals show substantially more pairwise divergence. This increased heterogeneity is the epigenomic correlate of increased phenotypic variability in aged tissues.

These three features together define what the authors term loss of epigenetic regulatory fidelity — a qualitative change in how methylation is organized and maintained, not merely a sum of site-specific gains and losses. This concept predates but is consistent with the Shangtong 2015 / Hannum/Esteller conceptual framework of epigenetic entropy increase with age. needs-replication — the specific spatial-autocorrelation metric has not been widely replicated in other tissues.

4. Age-related methylation differences — quantitative characteristics

The age-associated differentially methylated positions (DMPs) show:

• Predominantly small effect sizes (localized, modest β changes) rather than large global shifts.
• Both hyper- and hypomethylation, with discontinuous distribution — changes are not evenly spread across the genome but occur at specific regulatory elements.
• The set of age-correlated CpGs partially overlaps with known polycomb-target regions and bivalent chromatin domains (consistent with later findings in Qi 2026 ² confirming bivalent-region enrichment of age-correlated hypermethylation in epidermis across ethnicities).

5. Transcriptional connectivity

Gene expression analysis of paired epidermis samples (RNA-seq) shows:

• Young methylomes correspond to tightly interconnected transcriptional networks: a high fraction of gene pairs show statistically significant co-expression.
• Old methylomes show an increased number of gene pairs with no expression correlation — i.e., the transcriptional network becomes sparser and less coordinated.
• The erosion of methylation patterning and the reduction in transcriptional connectivity are correlated, supporting the interpretation that epigenetic disorganization drives transcriptional dysregulation rather than merely co-occurring with it. The causal direction (methylation → transcription vs. transcription → methylation) is not established in this paper. no-mechanism

Why this paper matters

1. Foundational tissue-specific clock

Bormann 2016 established that epidermis has an epigenetic aging signature measurable with substantially better precision than pan-tissue clocks. It created the methodological template for tissue-specific methylation clocks: collect homogeneous tissue, assay with 450k or equivalent, train SVM/elastic-net on age, validate in independent cohort. This template was followed by subsequent skin clocks including the Qi 2026 23k/173-CpG clock ² and the Rodríguez-Paredes 2026 TapeLift clock.

The Bormann 2016 training cohort (n=108, 18–78 yr) was reused by Qi 2026 as the foundational training set for the 23k clock, which used the intersection of the Bormann 27k/450k-shared 23,845 probes and selected 173 CpGs by penalised regression. The Bormann cohort thus anchors the entire skin epigenetic-clock literature.

2. Mechanistic insight — epigenetic entropy in epidermis

The spatial-autocorrelation and dynamic-range-erosion findings go beyond clock accuracy to characterize how the methylation landscape changes with age. This is the paper’s primary conceptual contribution: aging is not just the sum of site-specific gains and losses, but involves a structural reorganization of the methylation architecture (loss of domain boundaries, increased cell-to-cell heterogeneity). This is mechanistically distinct from the Horvath 2013 clock’s 353-CpG predictor, which is agnostic to the structural context of the CpGs selected.

The transcriptional-connectivity collapse finding links the methylation architecture change to functional consequences in gene regulation — providing a bridge between the epigenetic hallmark and the phenotypic outcomes of impaired tissue function in aged skin.

3. Benchmark for skin-intervention studies

The Vandiver comparison (Bormann MAE 6.72 yr vs Horvath MAE 12.1 yr) established that pan-tissue clocks systematically underperform in epidermal samples. This benchmark justified the development of the subsequent generation of skin-specific clocks and is the implicit reason Qi 2026 built a new clock rather than using Horvath.

Any intervention trial using epigenetic-age as a skin endpoint should use the Bormann 2016 / Qi 2026 / TapeLift generation of tissue-specific clocks rather than Horvath — the 12.1-yr MAE of Horvath in this tissue would swamp small intervention effects.

Strengths

• n=108 is the largest epidermis methylome dataset at time of publication (paper explicitly states this).
• Two-cohort design (punch biopsy + suction blister) provides sampling-method diversity; both biopsies and blister roofs yield viable epidermis.
• Independent validation in the Vandiver 2015 dataset (external, non-authors’ cohort) is methodologically appropriate.
• Multi-modal analysis (methylation clock + pattern analysis + transcriptomics) in the same cohort strengthens the convergent interpretation.
• Gold open access with PMC availability.
• Clock model reusability: the training cohort and probe set have been reused and extended by Qi 2026, creating a direct lineage.

Limitations

1. Female-only Caucasian cohort. All 108 donors are white females. The clock’s performance on males, non-Caucasian donors, or diverse phototypes is unknown from this paper. Qi 2026 subsequently showed that a new clock trained on the same Bormann 2016 cohort and validated in 17 multi-ethnic donors achieves similar MAE (4.88 yr) without ethnic bias — suggesting the underlying biology is ethnicity-conserved — but this was not tested in Bormann 2016 itself. needs-replication
2. Cross-sectional design. Like all cross-sectional methylation clocks, the predictor captures population-level age trends, not individual longitudinal trajectories. Whether the clock tracks within-person aging rate is not established.
3. Sun exposure not fully characterized. The validation cohort (Vandiver 2015) specifically labeled samples as “sun-exposed”; the training cohort used outer forearm samples, which receive moderate but variable UV exposure depending on individual behaviour. UV-exposed vs photoprotected skin comparisons are not performed within this paper. needs-replication
4. Model architecture limitations. SVM regression (as opposed to elastic-net / penalized regression) was state-of-the-art in 2016 but is less interpretable than the CpG-coefficient model of Horvath 2013. The SVM model is not easily transferred as a simple linear formula; practical deployment requires the full SVM parameters. Later tissue-specific clocks (Qi 2026 173-CpG elastic-net) addressed this portability limitation.
5. Causal direction of methylation ↔ transcription relationship. The co-occurrence of methylation-pattern erosion and transcriptional-connectivity loss is documented, but the causal direction is not established. Intervention or perturbation studies (e.g., DNMT depletion in keratinocytes + transcriptome readout) would be needed to establish directionality. no-mechanism
6. No mortality or morbidity outcomes. Like Horvath 2013, this is a chronological-age predictor only. Whether epidermal DNAm age acceleration (EAA) predicts wound-healing impairment, keratinocyte malignancy risk, or other clinically relevant skin outcomes is untested in this paper. needs-replication

Extrapolation Assessment

Dimension	Status	Notes
Pathway conserved in humans?	yes	This is a human study throughout; epidermis methylomes are human tissue
Findings replicated by independent groups?	partial	Clock concept extended by Qi 2026 + TapeLift clock; spatial-autocorrelation metric not widely replicated
Findings generalized beyond white females?	partial	Qi 2026 cross-ethnic validation of 23k clock derived from Bormann cohort; Bormann model itself not cross-validated in males or non-Caucasian samples

Cross-References

• skin-aging — phenotype page; Bormann 2016 is foundational for the § Epigenetic alterations — skin-specific epigenetic clocks section
• epidermis — tissue page; this study is the anchor citation for the epigenetic aging signature of the epidermis
• skin — skin tissue hub; links to Bormann 2016 for the skin-as-epigenetic-readout section
• epigenetic-alterations — Bormann’s methylation-pattern-erosion and transcriptional-connectivity findings are direct evidence for this hallmark
• horvath-clock-2013 — the pan-tissue clock that Bormann improves upon in epidermis-specific performance
• qi-2026-dhm-epigenetic-skin-aging — directly builds on Bormann 2016 training cohort; 23k clock is a redesigned version of the Bormann probe set
• bormann-epidermis-clock-2016 — R43 forward-ref for the planned biomarker page; this study page is the primary-source anchor for that page when it is seeded

Footnotes

Footnotes

1. horvath-2013-epigenetic-clock · n=7,844 samples / 51 tissues / 82 datasets · meta-analysis (cross-sectional) · Pearson r=0.96; MAE=3.6 yr held-out · model: multi-tissue pan-tissue 353-CpG clock · doi:10.1186/gb-2013-14-10-r115 · local PDF available; verified 2026-05-07 ↩

2. qi-2026-dhm-epigenetic-skin-aging · n=77 (17 pilot cross-ethnic + 60 product-use) · in-vivo observational (Part 1) + uncontrolled open-label (Part 2) · 23k clock CV MAE 5.66 yr; multi-ethnic validation MAE 4.88 yr; DHM-serum −2.1 yr DNAm age shift (paired Wilcoxon p=0.029) · model: human epidermis tape-strip (multi-ethnic) · doi:10.1007/s13555-026-01764-4 · verified: false (auto-extracted from open-access HTML; full PDF pending) ↩ ↩²