Do the Gene Signatures of Aging Reproduce?

The question

A now-familiar genre of headline promises to measure your "biological age" from a blood sample, and an industry sells tests, supplements, and regimens that claim to slow it. Underneath those claims is a research programme that looks for molecular signatures of aging: sets of genes whose activity rises or falls with age, or a statistical model that maps a gene-expression or DNA-methylation profile onto an estimated age. The scientific question this brief addresses is narrower and more basic.

When a study reports that a set of genes changes with age in one cohort, does that signature reproduce in independent cohorts and datasets? And when it fails to, why?

Reproducibility is the difference between a biomarker and a lucky fit. If an age-related gene signature is a stable feature of human biology, it can be used to probe mechanisms, stratify risk, and judge whether an intervention did anything. If it is partly an artifact of who was sampled, how the blood was handled, or which cell types happened to be present at the time, then a biological-age score built on it is measuring something other than what it claims. The evidence below concerns ordinary, general-audience biology; no clinical recommendation is offered.

Why it matters

Aging is the largest single risk factor for most chronic disease, and a biomarker that tracks the biological process would be valuable: for understanding mechanism, for prioritising drug targets, and for running shorter, cheaper trials of interventions that claim to delay decline. That promise has made "biological age" a consumer product before it is a validated clinical measurement.

The reproducibility question changes what any of it means. If the reproducible part of an "aging expression signature" is mostly the changing mix of immune cells in a blood draw, then a test that reports someone's biological age may be reporting their immune composition — informative, perhaps, but not the thing the buyer thinks is being measured, and not obviously responsive to an anti-aging drug. Conversely, if gene-level signatures are unstable but a smaller core is real, then effort should go to identifying that core rather than to ever-larger gene panels. The stakes are not exotic: they are the difference between a biomarker that guides research and one that sells certainty.

What counts as replication

Four different claims hide inside "the signature replicated," and they are not interchangeable.

• Same-data reproducibility. Re-running the same analysis on the same samples gives the same answer. This is a software check, not evidence about biology.
• Gene-level replication. Gene X's association with age — its direction and rough magnitude — reappears in a new, independent cohort.
• Signature-level replication. A model trained on one cohort keeps its predictive accuracy when applied to a new one.
• Mechanistic replication. The reason the genes move together is the same across cohorts.

A widely quoted high correlation between discovery and replication (a value of 0.972 in one landmark study [1]) is a same-study, same-protocol result, not independent replication. It shows the measurement is internally consistent, not that the signature travels. This distinction is the source of much of the confusion in the field.

Best current answer

The best evidence supports a qualified negative for gene expression, with an important qualifier. Age-associated transcriptomic signatures replicate markedly worse across independent cohorts than age-associated DNA-methylation signatures, and a large fraction of the reproducible transcriptomic signal is compositional.

The cleanest single piece of evidence is a head-to-head comparison. A 2026 study in Nature Communications analysed six large expression datasets and six methylation datasets, computing how consistently each top age-associated feature tracked age across cohorts [2]. Age-associated expression correlations were "quite variable between cohorts" and generally low: even the strongest aging transcripts changed their correlation with age from roughly −0.3 to −0.5 depending on the cohort. The canonical gene signature from the largest prior expression meta-analysis — 1,497 age-associated genes from 14,983 people [1] — also showed "generally low levels of correlation" when its genes were re-examined across independent datasets [2]. The same study's methylation features behaved differently: correlations with age were larger and more consistent across cohorts, with the same well-known site topping the list each time [2]. The authors concluded plainly that "age-related transcriptomic data appear to replicate poorly across cohorts."

That comparison should not be read as "aging leaves no trace in gene expression." It leaves a large trace. The problem is that the specific gene lists and model weights move around, and the part that stays put is largely the changing census of cell types in the sample (developed below). The units that replicate best are those engineered to be robust: a computationally integrated multi-omic score that combines methylation and expression, for instance, produced genes that "replicate more robustly across diverse populations" and survived adjustment for cell proportions for the majority of its members [2].

The composition problem

Most human aging transcriptomics has been done on whole blood, which is not one tissue but a suspension of cell types — neutrophils, lymphocytes, monocytes — each with its own transcriptome, in proportions that change with age. Two age-related shifts dominate: naive T cells decline, and the neutrophil-to-lymphocyte ratio rises.

The consequences are severe and quantified. A 2023 study in Genomics recruited 138 donors and measured both gene expression and leukocyte counts directly; it found that the combination of neutrophil and lymphocyte counts alone was the primary determinant of whole-blood transcript levels "for up to 75% of the protein-coding genes expressed in peripheral circulation," while age, sex, and common diseases had comparatively minor influence [5]. In other words, three-quarters of the transcriptome in a blood draw is partly a report on how many of which white cells were present.

The aging-gene lists reflect this. A 2024 analysis in Frontiers in Aging compared published whole-blood aging genes against gene-specificity data for leukocyte subtypes and found that the highly ranked aging genes "were predominantly expressed by naïve T cells," with limited expression from the more common cell types [3]. Its authors proposed that much of the observed expression change in aging blood reflects the decline of naive T cells "rather than changes in transcription rates in common cell types" — and warned that transcriptomic age models trained on unadjusted bulk blood "may be little more than predictors of naïve T cell decline" [3].

Adjustment changes the biology you recover. A 2021 study in BMC Genomics compared a model that corrected for the major cell types against one that also corrected for rarer subtypes across 3,165 blood samples; the extended correction "dramatically" altered effect sizes and significance, and the resulting age genes were markedly less tied to any single cell type [4]. The lesson is that partial correction leaves composition-driven false positives in place.

Composition also dwarfs the aging signal at the single-cell level. A 2024 integrated atlas pooled seven single-cell datasets — over a million cells from 103 donors — and re-annotated them consistently. When it compared old and young donors, the transcriptomic differences between age groups "were less pronounced than the variation between the datasets": hierarchical clustering grouped samples by dataset, not by cell type or age [8]. A few trends did replicate — declining CD8+ naive T cells and MAIT cells, expanding non-classical monocytes — but the dominant structure was between-study, not between-ages [8].

Mechanisms: why a signature fails to travel

At least four distinct mechanisms can make a reported signature fail to replicate, and they call for different fixes.

1. Biological composition. The signature captures a real but indirect effect — the changing proportions of cell types — that varies with the cohort's age structure, health, and sampling [3,5].
2. Technical and preprocessing artifacts. Platform, batch, RNA handling, and normalization choices move the numbers. A companion 2023 study in Scientific Reports showed that three common normalization strategies (median-ratio, trimmed-mean-of-M-values, and quantile normalization) "masked the true biological structure" of whole-blood data when donors differed in leukocyte counts; only simple scaling by sequencing depth preserved it [6].
3. Statistical overfitting. Cross-sectional designs with many genes and modest samples fit noise that does not generalise. A large 2021 meta-analysis in Aging Cell that reprocessed over 6,000 raw samples explicitly highlighted "the reproducibility limits of these machine-learning models" despite very high in-sample fit [7].
4. Real biological heterogeneity. Some aging changes are genuinely cell-type- and tissue-specific. A 2025 preprint building transcriptomic clocks across tissues reported that although all tissues shared a temporal "wave" pattern of gene importance, "each employs distinct gene sets to implement it" [18]. A 2025 study of human brain cell-type-specific clocks found that most clock feature genes were unique to a cell type, and that a signature's feature genes correlated with age in the same direction in only 40–80% of cases in one external dataset and 15–40% in another [12].

Mechanisms 1 and 2 are, in principle, fixable with better measurement; mechanism 3 with preregistration and held-out data; mechanism 4 is a fact about biology that limits how transportable any single signature can be.

Competing explanations

The field does not agree on which mechanism dominates, and the disagreement is substantive.

The methodological view. Signatures fail because the field has been sloppy — no cell correction, no harmonisation, no held-out validation. On this reading, the fix is better pipelines, and the reproducible results from careful studies support it. The strongest support is that when heterogeneity is modelled explicitly, signatures stabilise [11,13] and multi-omic integration recovers genes that replicate [2].

The intrinsic-heterogeneity view. Aging transcriptomes are genuinely diverse, and the whole notion of a single bulk "signature of aging" is a measurement error. On this reading, bulk reproducibility is limited by design, and the right unit of analysis is the cell type, not the person [8,12,18].

The immunosenescence view. There is a reproducible signal, but it is mostly immune-compositional — naive T-cell decline and myeloid skew — so what replicates is a measure of immune aging, not a general "biological age" [3,5,8]. This view is compatible with the first two: the signal is real, but narrower than advertised.

These are not mutually exclusive. The evidence best fits a combination: much early failure is methodological, the residue is partly intrinsic, and the reproducible core is heavily immune.

Counterevidence: where signatures do reproduce

The negative answer is not total, and the exceptions are instructive.

A 2026 study in GeroScience on 2,167 members of long-life families compared its whole-blood aging signature with the canonical Peters list and with an independent longevity cohort [11]. It reported significant replication: gene-level correlation of about 0.78 against the prior signature and 0.82 against the independent cohort, rising to about 0.94 when restricted to genes significant in either study, with a transcriptomic clock whose age acceleration predicted mortality (hazard ratio ≈ 1.11 per unit) [11]. The conditions matter: a large, deeply phenotyped cohort and a signature validated against external data, with strong immune involvement.

A 2026 study in BMC Bioinformatics took the opposite fix. It aggregated 16 public PBMC datasets (174 individuals) and built a Bayesian model that included the dataset as a random effect to absorb batch and technical heterogeneity [13]. The resulting predictor reached an R² of about 0.86 with a mean absolute error of 5.5 years, and converged on a "concise consensus signature of 70 genes, stable across datasets" pointing to natural-killer-cell biology [13]. Explicitly modelling study-level heterogeneity recovered stability — evidence for the methodological view.

The strongest single replication improvement came from adding a second molecular layer: integrating methylation with expression to find genes that change in both produced signatures that "much more robustly replicate across datasets" than expression-only signatures [2]. And where the goal is age prediction rather than mechanistic signature, a transcriptome clock in a controlled animal model has approached the theoretical accuracy limit [10], showing the ceiling is high when the data are homogeneous and the goal is narrow.

Evidence quality

The primary studies here are large and technically strong, but the evidence base has systematic limits. The samples are dominated by whole blood and by people of European ancestry; several key analyses are cross-sectional rather than longitudinal; and few published signatures were validated prospectively on held-out cohorts before being described as reproducible. Because there is no gold-standard molecular definition of "biological age," accuracy is usually judged against chronological age or future health, which rewards signals that track time and composition rather than causal aging.

A useful calibration comes from the more mature DNA-methylation clock literature. Even there, the same confounds recur: cell composition explains up to about 53% of the variation in methylation age [14]; between roughly 24% and 54% of clock sites in major models are "incoherent," with coefficients pointing opposite to their univariate trend with age [15]; and although technical reproducibility across replicates is excellent, biological reliability across meals, stress, and repeated sampling is only moderate, with most clocks falling short of clinical stability [16]. That DNA-methylation signatures replicate better than expression signatures, while still carrying substantial composition and reliability problems, sets the bar the expression literature has not yet cleared. Two of the cited works are preprints and are not peer-reviewed [16,18].

Limitations of this brief

This brief cannot put a precise number on what fraction of aging genes replicate in held-out cohorts — no single preregistered benchmark has computed it. It compares gene expression mainly against DNA methylation, not against proteomics or metabolomics. It is restricted to human bulk and single-cell transcriptomics with limited cross-species coverage, and the mechanistic "wave" claim [18] rests on a preprint. And a bounded search cannot rule out a well-powered study that would revise the picture; the honest statement is that the evidence supports the direction, not an exact magnitude.

Calibrated confidence

• ~85% that age-associated transcriptomic signatures replicate substantially worse across independent cohorts than age-associated DNA-methylation signatures [2].
• ~75% that a large share of the reproducible whole-blood transcriptomic age signal reflects changes in leukocyte composition — naive T-cell decline and the neutrophil-to-lymphocyte ratio — rather than cell-intrinsic transcription [3,5].
• ~65% that cell-type-resolved signatures are more reproducible than bulk signatures but remain dataset-sensitive [8,12].
• ~60% that the reproducible core of aging expression is a relatively small, immune-dominated set that generalises when study-level heterogeneity and composition are modelled explicitly [11,13].
• ~80% that the field still lacks a preregistered, held-out, multi-cohort benchmark quantifying how much of a frozen signature replicates after full cell-type deconvolution.

Reversal conditions

This brief's central claim would weaken or reverse if:

1. A large, harmonised, preregistered multi-cohort benchmark applied a frozen bulk expression signature to independent cohorts and found high gene-level replication after full cell-type deconvolution — showing the earlier failures were mostly methodological.
2. A signature-level model retained accuracy across cohorts, species, and tissues without cell-composition adjustment or a second omic layer.
3. The variance currently attributed to cell composition were shown to be largely cell-intrinsic change that co-varies with composition, undercutting the distinction between "who is in the sample" and "how the cells are aging."

Each is a testable event, and any of them would move the confidence estimates materially.

The gap that would settle it

The literature gives a direction (expression signatures replicate worse and are composition-heavy) but not a magnitude. The exact unresolved claim is: for a fixed, published aging transcriptomic signature, what fraction of its genes replicate in held-out independent cohorts at each level of adjustment — none, bulk cell-count correction, and single-cell deconvolution — and how much of the reproducible fraction is cell-intrinsic?

That is a bounded project candidate rather than a synthesis, because producing the number requires an original reanalysis of existing public data. It does not require new data collection. First feasibility check: determine whether enough public bulk RNA-seq cohorts with gene-level data, wide age ranges, and either matched cell-count measurements or single-cell references exist (for example in GEO/ArrayExpress or aggregated repositories such as ARCHS4), and whether current deconvolution tools can be run consistently across them. If the cohorts or references are insufficient, the claim stays a documented gap; if they exist, the project can register a frozen signature and a held-out replication protocol before opening the confirmation data.

References

1. Peters MJ, Joehanes R, Pilling LC, et al. The transcriptional landscape of age in human peripheral blood. Nature Communications. 2015;6:8570. doi:10.1038/ncomms9570.
2. Moqri M, Ying K, Poganik JR, et al. Integrative epigenetics and transcriptomics identify aging genes in human blood. Nature Communications. 2026;17:725. doi:10.1038/s41467-025-67369-1.
3. Fraser C, Owen BM. Naïve T-cell decline is a significant contributor to expression changes in ageing blood. Frontiers in Aging. 2024;5:1389789. doi:10.3389/fragi.2024.1389789.
4. Pellegrino-Coppola D, Claringbould A, Stutvoet M, et al. Correction for both common and rare cell types in blood is important to identify genes that correlate with age. BMC Genomics. 2021;22:184. doi:10.1186/s12864-020-07344-w.
5. O'Connell GC, Wang J, Smothers CG. Donor white blood cell differential is the single largest determinant of whole blood gene expression patterns. Genomics. 2023;115(6):110708. doi:10.1016/j.ygeno.2023.110708.
6. O'Connell GC, et al. Variability in donor leukocyte counts confound the use of common RNA sequencing data normalization strategies in transcriptomic biomarker studies performed with whole blood. Scientific Reports. 2023;13:15514. doi:10.1038/s41598-023-41443-4.
7. Shokhirev MN, Johnson AA. Modeling the human aging transcriptome across tissues, health status, and sex. Aging Cell. 2021;20(1):e13280. doi:10.1111/acel.13280.
8. Filippov I, Schauser L, Peterson P. An integrated single-cell atlas of blood immune cells in aging. npj Aging. 2024;10:59. doi:10.1038/s41514-024-00185-x.
9. de Magalhães JP, Curado J, Church GM. Meta-analysis of age-related gene expression profiles identifies common signatures of aging. Bioinformatics. 2009;25(7):875–881. doi:10.1093/bioinformatics/btp073.
10. Meyer DH, Schumacher B. BiT age: A transcriptome-based aging clock near the theoretical limit of accuracy. Aging Cell. 2021;20(3):e13320. doi:10.1111/acel.13320.
11. Li M, Song Z, Reed ER, et al. Whole blood transcriptional signatures of age and survival identified in long life family and integrative longevity omics studies. GeroScience. 2026;48(6):8621–8639. doi:10.1007/s11357-025-02090-x.
12. Muralidharan C, Zakar-Polyák E, Adami A, et al. Human brain cell-type-specific aging clocks based on single-nuclei transcriptomics. Advanced Science. 2025;12(43):e06109. doi:10.1002/advs.202506109.
13. Suaste V, Daza-Torres ML, Montesinos-López JC, et al. BAGE: a Bayesian framework for age prediction based on PBMC gene expression data. BMC Bioinformatics. 2026;27:193. doi:10.1186/s12859-026-06519-8.
14. Jonkman TH, Richmond A, Marioni RE, van Zwet EW, Heijmans BT. Blood cell composition reveals distinct biological interpretation of DNA methylation age and age acceleration. Genome Medicine. 2026;18:142. doi:10.1186/s13073-026-01751-6.
15. Mei X, Kabir H, Conboy MJ, Conboy IM. Misalignment of age clocks. GeroScience. 2025;47(5):6647–6658. doi:10.1007/s11357-025-01750-2.
16. Sehgal R, Borrus D, Gonzalez J, et al. Biological versus technical reliability of epigenetic clocks and implications for disease prognosis and intervention response. bioRxiv preprint. 2025. doi:10.1101/2025.10.13.682176.
17. Arpawong TE, Cole S, Badhesha H, et al. How epigenetic clocks tick: unpacking the black box by deciphering biological pathways and transcriptomic signatures of accelerated aging. npj Aging. 2026. doi:10.1038/s41514-026-00446-x.
18. Agarwal V, Li O, Petty CA, et al. Guided multi-agent AI invents highly accurate, uncertainty-aware transcriptomic aging clocks. bioRxiv preprint. 2025. doi:10.1101/2025.09.08.674588.