https://www.nature.com/articles/s41586-026-10542-3
chatGPT:
Summary
The paper, “Universal transcriptomic hallmarks of mammalian ageing and mortality,” builds large-scale RNA-expression clocks for ageing and mortality across mammals. The authors integrate more than 11,000 transcriptomes, from more than 25 tissues and four mammals: mouse, rat, macaque and human. Their aim is not just to predict chronological age, but to identify transcriptomic patterns linked to mortality risk, lifespan-modulating interventions, disease, damage accumulation and rejuvenation.
The core idea is that transcriptomic ageing is partly conserved across tissues, species and cell types. The authors train rodent and multi-species clocks for:
- Chronological age
- Normalized age, meaning age adjusted for expected species/strain/intervention lifespan
- Expected mortality, estimated using survival models such as Gompertz fits
In rodents, they use data from UM-HET3 mice exposed to 20 Interventions Testing Program treatments, plus a broader rodent meta-dataset of 4,539 transcriptomes across 26 tissues and 79 interventions. They report that relative transcriptomic clocks perform well: for example, the rodent relative chronological clock achieved about r = 0.96, and the rodent mortality clock about r = 0.95 in held-out data.
A major finding is that many age-associated transcriptomic changes seem to track health deterioration rather than neutral ageing drift. Inflammatory, interferon, p53, complement and coagulation pathways tend to rise with ageing and mortality, whereas oxidative phosphorylation, mitochondrial translation, fatty acid metabolism and xenobiotic metabolism tend to associate with longer lifespan or lower mortality.
The authors then extend the framework across species. They add 2,623 macaque samples and 4,003 human samples, producing multi-species “universal” transcriptomic clocks. Conserved age-upregulated genes include GPNMB, VSIG4, CDKN1A and EDA2R, while genes such as NREP, COL1A1 and COL3A1 decline with age. They argue that this supports a conserved mammalian transcriptomic architecture of ageing.
A particularly interesting part is the module-specific clocks. Instead of treating ageing as one global score, they identify co-expression modules corresponding to inflammation, interferon signalling, mitochondrial function, chromatin modification, extracellular matrix organisation, mRNA splicing, metabolism and other systems. These module clocks reveal that different interventions or diseases affect different biological subsystems. For example, chronic diseases primarily accelerate inflammatory-module ageing, whereas caloric restriction and Klotho deficiency mainly affect metabolic and mitochondrial modules.
The paper also connects transcriptomic clocks to human outcomes. It reports that transcriptomic and DNA methylation age acceleration correlate in human blood, with the strongest relationship involving a chromatin-associated transcriptomic module. It also highlights proteins such as CDKN1A and LGALS3, whose plasma levels are associated with mortality and multimorbidity in UK Biobank.
Novelty
The main novelty is not simply another ageing clock. The paper combines several advances:
First, it shifts from chronological-age prediction to mortality-associated transcriptomic ageing. The mortality clock is designed to capture not just how old an organism is, but how its expression profile relates to expected hazard and lifespan modulation.
Second, the study integrates ageing, lifespan-shortening models and lifespan-extending interventions in one framework. The authors explicitly note that previous work had not provided a unified analysis of mortality-associated mechanisms shared across ageing, short-lived models and longevity interventions.
Third, the work is cross-species and multi-tissue. It attempts to find transcriptomic signatures that persist across rodents and primates, rather than being restricted to one tissue or one organism.
Fourth, the module-clock approach is valuable. A single “biological age” number can obscure mechanism. Here, the authors split ageing into subsystem-level clocks, making it possible to say, for example, that an intervention appears to affect inflammatory ageing more than mitochondrial ageing, or vice versa.
Fifth, the paper tries to bridge bulk RNA-seq, single-cell RNA-seq, disease datasets, rejuvenation models, DNA methylation clocks and plasma proteomics. That breadth makes the paper more of a systems-level ageing atlas than a narrow biomarker paper.
Critique
The study is impressive in scale and integration, but there are several important limitations.
The biggest conceptual issue is that mortality is inferred rather than directly measured for most transcriptomic samples. The authors estimate expected mortality using survival curves and Gompertz models for cohorts, strains, sex and interventions. That is reasonable, but the clock is partly trained on model-derived hazard labels, not individual observed death outcomes. This means the mortality clock may learn the structure of the survival modelling assumptions as well as biology.
A second issue is confounding by inflammation and disease burden. Many of the strongest mortality-associated signals are inflammatory: interferon, complement, p53, interleukin and innate immune pathways. These are plausible ageing mechanisms, but they are also generic responses to infection, tissue injury, immune activation and chronic disease. The paper does use module clocks to distinguish pathways, but a raised inflammatory tAge may still be a broad “sickness/injury” marker rather than a specific ageing mechanism.
Third, the use of relative expression centring helps reduce batch effects, but it also makes the clocks dependent on the choice of reference controls. This is practical for controlled datasets, but harder in clinical use. A transcriptomic age calculator may work well when samples are processed in a matched study design, but performance may degrade when applied to heterogeneous real-world data.
Fourth, although transcriptomic clocks are more interpretable than DNA methylation clocks, expression is highly state-dependent. RNA profiles change with circadian timing, diet, acute stress, infection, medication, cell composition, tissue sampling method and batch effects. That makes transcriptomic age potentially more responsive, but also more fragile as a stable ageing measure.
Fifth, the paper risks treating correlation as mechanism. Genes such as CDKN1A, LGALS3 and GPNMB may be excellent biomarkers of ageing or mortality, but that does not prove they are causal drivers. CDKN1A/p21, for instance, may reflect senescence, DNA damage response or stress signalling, but whether lowering it would improve ageing outcomes depends heavily on context, including cancer suppression.
Sixth, the “universal” claim should be interpreted cautiously. The paper covers four mammals and many tissues, which is strong, but that is still a limited slice of mammalian diversity. Humans are represented mainly through available tissues such as blood, brain, skin and muscle, not full organism-wide sampling.
Finally, the module clocks are biologically attractive, but their individual predictive accuracy is more modest than the global clocks. That is expected, because each module contains fewer genes, but it means module-level interpretation should be used as a hypothesis-generating tool rather than as definitive pathway diagnosis.
Bottom line
This is a strong and ambitious paper. Its most important contribution is a multi-species, multi-tissue transcriptomic framework that links ageing, mortality, lifespan interventions, chronic disease and rejuvenation-like states. The module-specific clocks are especially useful because they move the field away from a single biological-age number toward a more mechanistic map of ageing subsystems.
The main caution is that the clocks are still biomarkers, not proof of causal mechanisms. They appear powerful for comparing interventions and disease states, but their interpretation depends heavily on sampling context, inflammation, batch correction and the survival models used to define expected mortality.