Transcriptomic clock (tAge)

A transcriptomic clock estimates biological age from a tissue’s (or cell’s, or pseudobulk metacell’s) gene-expression profile. Unlike DNA-methylation clocks (which read a slow, largely stochastic epigenetic mark) transcriptomic clocks read the dynamic functional state of a sample, making them sensitive to acute damage and rejuvenation that DNAm misses. The defining work is Tyshkovskiy et al. 2026 (Gladyshev lab) ¹, which built interpretable, mortality-trained, multi-species transcriptomic clocks and is the canonical reference for this page.

This page covers the tAge clock family as a unit (they share training data, software and a single primary reference), rather than splitting per sub-clock.

The clock family

Clock	Target	What it captures
Chronological tAge	calendar age (scaled to species max lifespan)	age per se; largely insensitive to lifespan-extending interventions
Normalized-age tAge	chronological age ÷ expected max (99.9th-pct) lifespan	age and intervention effects
Mortality tAge	expected Gompertz log-hazard rate	the flagship: ageing + intervention + damage burden; best lifespan-model discriminator
Maximum-lifespan (“lifespan”) clock	species/strain max lifespan	longevity-regulating expression, partly age-independent
Module-specific clocks (23 rodent / 14 multi-species)	per WGCNA module	which subsystem is ageing — inflammation, interferon, OXPHOS, chromatin, ECM/EMT, …

The mortality clock is the key conceptual advance. Chronological clocks barely move under rapamycin/CR-type interventions, so they cannot grade an intervention’s benefit. The mortality clock — trained on cohort-specific Gompertz expected-hazard rather than calendar age — integrates ageing-associated and intervention-modulating changes, and discriminates short- vs long-lived models better than even a dedicated lifespan clock. This is the transcriptomic analogue of the leap from first-generation DNAm clocks (horvath-clock-2013, hannum-clock-2013) to mortality-trained second-generation clocks (grimage-2019, phenoage-2018, dunedinpace-2022).

How it is built

• Models: elastic-net (sparse, interpretable gene coefficients) and Bayesian-ridge (probabilistic, uncertainty-quantified — useful for small samples and single-cell). Elastic-net and BR outperformed SVM, random forest, kNN and LightGBM.
• Relative (“reference-centred”) clocks: each sample’s expression is centred against a randomly chosen age/sex/strain-matched control reference group within the same dataset+tissue, then the clock predicts the age difference. This within-dataset batch correction lifted rodent multi-tissue accuracy to median R²=0.92 / r=0.96 and was robust to reference-group choice (LOFO r=0.939–0.956).
• Mortality target: per-cohort/strain/sex/intervention Gompertz fits (μ(t)=Ae^{rt}) give each sample an expected log-hazard; normalized age = chronological age ÷ Gompertz-simulated 99.9th-percentile max lifespan.
• Species scaling: chronological age divided by species maximum recorded lifespan (4 / 3.8 / 39 / 122 yr for mouse / rat / crab-eating macaque / human) to make a unified multi-species clock.
• Module clocks: WGCNA co-expression modules (28 rodent / 15 multi-species) → elastic-net clock per module; the composite clock decomposes additively into per-module contributions (tAge_composite = C₀ + Σ tAge_module).
• Single-cell application: bulk-trained clocks are applied to snRNA-seq data via metacell aggregation (pooling ~100+ cells / ≥200k reads), which bridges single-cell sparsity to clock input — used for the Tabula Muris Senis and Klotho-KO analyses.

Performance

Setting	Metric
Rodent chronological (multi-tissue)	R²=0.88, r=0.94 (held-out)
Rodent relative chronological	median R²=0.92, r=0.96
Multi-species chronological (LOFO, 4 species)	r=0.952 — matches pan-mammalian DNAm clock (r=0.953) 2
Multi-species mortality clock	r=0.91 (macaque) / 0.94 (human) / 0.96 (mouse) / 0.92 (rat)
Human blood time-to-death (Framingham, RNA-seq n=3,698)	comparable to 2nd-gen DNAm clocks (DunedinPACE, YingDamAge, PhenoAge); beats 1st-gen transcriptomic chronological clocks (Peters, RNAAgeCalc)

What raises and lowers tAge (validation panel)

Raises mortality tAge (pro-ageing/damage):

• Replicative senescence in human fibroblasts + WI-38 (tAge rises with passage, precedes phenotype; CDKN1A a top driver)
• γ-irradiation (20 Gy) in mouse + naked-mole-rat fibroblasts; oligomycin and 2-deoxyglucose (metabolic poisons)
• Chronic disease (AD/5xFAD, CKD, stroke, NASH, diabetic nephropathy) — via inflammatory modules
• Klotho-KO progeria — via energy-metabolism + NRF2 modules (klotho)
• LPS neuroinflammation

Lowers tAge (rejuvenation):

• hTERT immortalization (abolishes the senescence-associated tAge rise)
• iPSC/partial reprogramming (in-vivo-partial-reprogramming-therapy) — strongest in EMT/MET module
• Heterochronic parabiosis in old mice (top driver Nrep; Cdkn1a/Vcam1 down)
• Caloric restriction (caloric-restriction) — via metabolic modules
• Early embryogenesis — U-shaped trajectory, minimum (“ground zero”) ~E10 (information-theory-of-aging)

Transcriptomic vs DNA-methylation clocks (complementary modalities)

	DNAm clocks	Transcriptomic clocks
Reads	stable epigenetic mark	dynamic functional state
γ-irradiation / hTERT	no change	tAge moves
Mechanistic interpretability	low (CpG → gene unclear)	high (genes annotated, module-resolved)
Single-cell application	limited	works (Bayesian-ridge on metacells)

In Framingham, transcriptomic and DNAm clock outputs were positively correlated after age/sex adjustment, strongest between the chromatin-modification transcriptomic module clock and epigenetic clocks — a candidate mechanistic bridge between the two modalities. The two are best read as complementary: DNAm = long-term stochastic drift, transcriptome = current damage/repair state.

Top contributing genes

Up with mortality tAge: CDKN1A, GPNMB, LGALS3, Cst7, Eda2r, Vsig4, S100a8, S100a6, Casp1. Down: Sparc, Nrep, Col1a1, Col3a1. See tyshkovskiy-2026-universal-transcriptomic-hallmarks for the full panel and module assignments.

Resources

• TACO (Transcriptomic Age Calculator Online): https://app.gladyshevlab.org/TACO/ — upload expression data, choose organ-specific / multi-tissue / multi-species clocks, compare groups.
• tAge R package; trained weights on Zenodo (doi:10.5281/zenodo.18763485).

Gaps & caveats

• Mortality-clock training inherits the quality of per-cohort Gompertz survival fits.
• Human intervention-responsiveness untested in an RCT — the open analogue of CALERIE-2/DunedinPACE. needs-human-replication
• Earlier transcriptomic clocks (Peters 2015 blood, RNAAgeCalc, Buckley 2023 brain) were chronological-only and human-restricted; this family is the first multi-species, mortality-trained, module-decomposed transcriptomic clock. Those predecessors are not yet separate wiki pages. needs-replication

Related pages

• Source: tyshkovskiy-2026-universal-transcriptomic-hallmarks
• DNAm clocks: horvath-clock-2013 · hannum-clock-2013 · phenoage-2018 · grimage-2019 · dunedinpace-2022
• Proteomic clock: lehallier-proteomic-clock-2019
• Driver genes: gpnmb · lgals3 · p21

Footnotes

1. tyshkovskiy-2026-universal-transcriptomic-hallmarks · n=11,165 transcriptomes, 4 species · meta-analysis + new in-vivo RNA-seq · model: mouse/rat/macaque/human ↩

2. Benchmark = Lu et al. 2023 pan-mammalian DNA-methylation clock (cross-species r=0.953); the multi-species transcriptomic chronological clock reached r=0.952. · model: cross-species ↩