Aging Biomarkers Compared: What Each Measure Captures—and What It Cannot Prove

1. Evidence summary

Aging biomarkers are not one technology or one biological construct. The label covers models trained to predict chronological age, scores associated with later disease or mortality, measures intended to estimate a pace of change, organ-specific models, functional tests and composites built from routine clinical inputs. Their outputs are not interchangeable. (Sources: Moqri et al., Validation of biomarkers of aging, PMID 38355974; Horvath 2013 multi-tissue clock, PMID 24138928.)

Epigenetic clocks are the most visible category, but the field now includes proteomic, metabolomic, immune, inflammatory and organ-specific approaches. Several have shown prospective associations with morbidity, function or mortality in studied cohorts. Those associations can make a measure useful for research or risk modelling without making it a clinical diagnosis, a direct measurement of aging rate or a validated surrogate endpoint. (Sources: Levine et al. DNAm PhenoAge, PMCID PMC5940111; Lu et al. DNAm GrimAge, PMID 30669119.)

Functional measures such as grip strength, gait speed and chair-rise performance differ from molecular clocks. They capture integrated capacity rather than a molecular mechanism, and they often predict adverse outcomes across older populations. Their prognostic value does not make them specific to aging biology: disease, injury, motivation, disability and environment also shape performance. (Sources: Expert consensus on biomarkers for intervention studies, PMID 39708300; Physical capability and mortality meta-analysis, PMID 20829298.)

The controlling boundary is crucial. A score may predict age, associate with risk or move after an intervention without demonstrating that the intervention changed the rate of aging, improved healthspan or will improve a clinical outcome. Surrogate validation requires evidence that change in the marker reliably predicts clinical benefit in a defined context. No aging-wide surrogate currently meets that standard. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

2. What an aging biomarker is

A biomarker is a defined characteristic measured as an indicator of biological processes, pathogenic processes or responses to an exposure or intervention. That broad definition describes what is measured; it does not determine how the result may be used. A measure intended for descriptive research faces a different validation burden from one used for prognosis, treatment selection or as a substitute for a clinical endpoint. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

In aging research, the word “age” can obscure these distinctions. A model may output a number in years because it was trained against chronological age. Another may output a risk-weighted “age” because its training target included mortality or clinical chemistry. A third may estimate the pace of longitudinal physiological change. Similar-looking numbers can therefore represent different targets, inputs and assumptions. (Sources: Horvath 2013 multi-tissue clock, PMID 24138928; Belsky et al. comparison of 11 measures, PMID 29149257.)

Professional interpretation starts with the intended use: What was the model trained to predict? In whom was it validated? How repeatable is the assay and calculation? Does it add information beyond ordinary risk factors? Does change track an outcome that matters? Without those questions, “biological age” can become a marketing label rather than a defined measurement claim. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

3. Prediction, association, responsiveness and surrogacy

Chronological-age prediction asks whether a pattern can reproduce calendar age. It is a model-performance question, not evidence that the output captures the causal rate of aging. A clock can be highly accurate at age prediction while being less informative about future disease, intervention response or individual clinical decisions. (Sources: Horvath 2013 multi-tissue clock, PMID 24138928.)

Outcome association asks whether a measure is related to later morbidity, disability or mortality. Outcome-trained epigenetic clocks, proteomic scores, metabolomic measures and functional tests can show such relationships. Association can support prognostic research, but residual confounding, disease burden and population differences limit causal interpretation. (Sources: Levine et al. DNAm PhenoAge, PMCID PMC5940111; Lu et al. DNAm GrimAge, PMID 30669119.)

Intervention responsiveness asks whether the measure changes after an exposure or intervention. That can be useful for detecting biological response. It is not the same as showing that the change mediates or predicts clinical benefit. A score could move because of assay variation, shifts in cell composition, acute metabolic effects or a process unrelated to long-term outcomes. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

Surrogacy is the strongest and most frequently overextended claim. A validated surrogate substitutes for a clinical endpoint only in a defined context supported by evidence connecting intervention-induced changes to clinical benefit. Responsiveness alone cannot supply that link. “Clock reversal” is therefore not proof of rejuvenation. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

4. Epigenetic clocks

First-generation epigenetic clocks use DNA-methylation patterns to predict chronological age across one or more tissues. Their achievement was analytical: a relatively small set of methylation sites could reconstruct age with substantial accuracy. The output is appropriately described as predicted age or age deviation, not a direct measurement of cellular aging speed. (Sources: Horvath 2013 multi-tissue clock, PMID 24138928.)

Later clocks were trained against composite phenotypes or mortality-related targets. PhenoAge and GrimAge illustrate this outcome-oriented shift. In validation cohorts, such measures have been associated with morbidity or mortality beyond chronological age. Because their targets incorporate clinical or exposure-related information, their interpretation differs from a pure chronological clock. (Sources: Levine et al. DNAm PhenoAge, PMCID PMC5940111; Lu et al. DNAm GrimAge, PMID 30669119.)

DunedinPACE takes another approach: its training target was longitudinal change across multiple organ-system indicators in a birth cohort. It is designed as a pace score rather than an age-in-years estimate. That design is closer to the question of rate, but it still depends on its training cohort, methylation assay, algorithm and external validation. It is not a direct stopwatch for aging. (Sources: Belsky et al. DunedinPACE, PMID 35029144.)

Across these families, blood-cell composition, tissue choice, preprocessing, assay platform and technical batch effects can influence results. Different clocks can disagree because they were built to capture different targets. Discordance is not necessarily an error, but it prevents the unqualified claim that one result is a person’s “true” biological age. (Sources: Belsky et al. comparison of 11 measures, PMID 29149257.)

5. Proteomic measures

Proteomic clocks and signatures use concentrations or patterns of circulating proteins. Large biobank studies have reported models that predict age or associate with incident disease and mortality, with some external validation. Proteins may reflect distributed signals across inflammation, metabolism, tissue injury and organ function, making them potentially informative but not mechanistically specific. (Sources: Kuo et al. proteomic aging clock, PMID 38747160; Tanaka et al. plasma proteomic signature, PMID 33210602.)

Platform dependence is a central constraint. Different affinity-based or mass-spectrometry technologies measure different protein sets with different calibration and batch properties. A model validated on one platform cannot be assumed to transfer unchanged to another. Disease and medication can also alter the same proteins used in an “aging” signature. (Sources: Kuo et al. proteomic aging clock, PMID 38747160; Tanaka et al. plasma proteomic signature, PMID 33210602.)

Proteomic scores remain primarily research tools. Prospective association is promising, but clinical actionability requires evidence that the score changes decisions or outcomes beyond established assessment. A risk-associated proteomic age gap is not a diagnosis of accelerated aging. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

6. Metabolomic measures

Metabolomic age models combine small molecules measured in plasma or serum into an age or risk score. The MileAge study reported associations between metabolomic age deviation and outcomes in UK Biobank, including an elevated hazard reported for an older-than-chronological grouping. The result supports association in that dataset, not a causal aging measurement. (Sources: MileAge metabolomic clock, PMID 39693428.)

Metabolites are sensitive to fasting state, diet, medication, kidney and liver function, disease and sample handling. That sensitivity can be useful when the goal is to capture current biological state, but it complicates interpretation as a stable intrinsic rate of aging. Reproducibility therefore depends on pre-analytic controls and platform consistency. (Sources: MileAge metabolomic clock, PMID 39693428.)

As with proteomics, the route to clinical use requires more than predictive performance. External validation, incremental utility, decision consequences and intervention-linked outcomes remain necessary. A metabolomic score should not be translated into a consumer instruction or treatment target from association alone. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

7. Immune and inflammatory measures

Immune-age and inflammatory-age models use cytokines, immune-cell distributions, transcriptomic data or multi-omic combinations. The iAge work linked an immune-aging score and CXCL9-related biology to multimorbidity, frailty and vascular or immune features. It is an informative research model, not a universal immune-age test. (Sources: Sayed et al. iAge, PMID 34888528.)

Immune measurements are especially context sensitive. Infection, vaccination, chronic disease, medication, adiposity and cell-composition shifts can affect inputs. Different immune clocks may capture different axes of immune state, and cross-platform agreement remains limited. (Sources: Belsky et al. comparison of 11 measures, PMID 29149257; Sayed et al. iAge, PMID 34888528.)

These measures may help organize research into immune remodeling and age-related risk. They cannot establish that a specific immune intervention will slow aging, or that a single high or low value supplies an individual diagnosis. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

8. Organ-age models

Organ-age models use organ-enriched plasma proteins, imaging or other inputs to estimate organ-specific deviations from expected age. One multi-cohort study reported that organ-age gaps were associated with organ-specific disease and mortality outcomes. The concept may expose heterogeneity hidden by a single whole-body score. (Sources: Oh et al. organ-aging proteomics, PMID 38057571.)

The output is nevertheless model defined. “Older heart” or “older kidney” in such a model does not mean that the organ’s literal age has been measured, nor does it predict which organ will fail in an individual. Replication, calibration, prospective utility and evidence that acting on the result improves care remain incomplete. (Sources: Oh et al. organ-aging proteomics, PMID 38057571.)

Organ-specific models should therefore be described as emerging research signatures. Their strongest current value is hypothesis generation and risk research, not individualized treatment selection. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

9. Functional measures

Grip strength, gait speed, chair-rise performance, balance and cognitive testing measure integrated capacity rather than a molecular pathway. Meta-analyses and consensus work support associations between physical capability and later adverse outcomes, including mortality. Standardized protocols can produce good reproducibility at low cost. (Sources: Expert consensus on biomarkers for intervention studies, PMID 39708300; Physical capability and mortality meta-analysis, PMID 20829298.)

Functional measures can also respond to rehabilitation, training, acute illness or environmental change. That makes them clinically tangible, but not specific to aging mechanisms. A decline can reflect disease or injury, while improvement can occur without demonstrating a change in fundamental aging biology. (Sources: Expert consensus on biomarkers for intervention studies, PMID 39708300; Physical capability and mortality meta-analysis, PMID 20829298.)

Their practical strength is that they measure capabilities that matter. Their limitation is specificity. They should not be dismissed because they are not molecular, nor relabelled as direct aging-rate measures because they predict outcomes. (Sources: Expert consensus on biomarkers for intervention studies, PMID 39708300; Physical capability and mortality meta-analysis, PMID 20829298.)

10. Composite clinical panels

Composite clinical-age models combine routine laboratory values, vital signs or functional inputs into a score. Their inputs may already be standardized and available, and some models show prospective risk associations. They can therefore offer an efficient summary of multisystem clinical state. (Sources: Levine et al. DNAm PhenoAge, PMCID PMC5940111; Belsky et al. comparison of 11 measures, PMID 29149257.)

The same convenience can conceal construct ambiguity. A composite may restate known cardiovascular, metabolic, kidney or inflammatory risk rather than measure an independent aging mechanism. If its output improves, the change may reflect one input or short-term physiological shift. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

Clinical availability of the inputs does not validate the composite’s biological-age claim. The complete model, calibration, population and intended decision need independent evaluation. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

11. Cross-platform comparison matrix

No row is designated “best.” Selection must follow the intended claim and validation context.

12. Reproducibility and population transfer

Reproducibility includes laboratory precision, pre-analytic control, algorithm stability and repeatability over relevant time intervals. A technically repeatable result can still be biologically ambiguous, while a biologically meaningful signal can be difficult to reproduce if sampling and processing differ. (Sources: Moqri et al., Validation of biomarkers of aging, PMID 38355974; Belsky et al. comparison of 11 measures, PMID 29149257.)

Population transfer is a separate test. Models trained in a particular age range, ancestry composition, health profile or healthcare setting may miscalibrate elsewhere. External validation should examine both rank ordering and absolute calibration, not merely report a statistically significant association. (Sources: Moqri et al., Validation of biomarkers of aging, PMID 38355974; Belsky et al. comparison of 11 measures, PMID 29149257.)

Commercial implementation can introduce additional changes in assay platform or calculation. A product sharing a scientific category with a published model is not automatically the same validated measure. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

13. Intervention responsiveness

Intervention studies often report that a clock or score changed after treatment, behavior change or altered exposure. Such findings can be legitimate exploratory evidence of biological response. Their meaning depends on prespecification, control group, assay precision, multiple-testing correction, effect size, durability and relationship to clinical outcomes. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

A favorable direction is not self-interpreting. If a measure was trained against chronological age, a small shift need not equal a corresponding number of years gained. If it was trained against risk, the intervention-induced change need not reproduce the risk relationship observed between people. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

The responsible wording is that a measure changed under specified conditions. Claims of rejuvenation, aging reversal or improved healthspan require direct outcome evidence or validated surrogacy that the change alone does not supply. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

14. Clinical actionability

Clinical actionability asks whether a result changes a decision in a way shown to improve outcomes. Most aging clocks and multi-omic signatures do not yet have that evidence. Risk-associated inputs may still identify ordinary clinical issues, but the actionable response generally comes from established disease or risk pathways rather than the “biological age” label. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

Functional measures and routine clinical panels are closer to current practice because their components can inform established assessment. Even there, the composite aging interpretation should not replace evaluation of the underlying findings. (Sources: Expert consensus on biomarkers for intervention studies, PMID 39708300; Physical capability and mortality meta-analysis, PMID 20829298.)

This record therefore does not rank consumer tests or explain individual results. Its purpose is to match each measure to the claim it can safely support.

15. What this establishes

• Aging measures are a family of distinct constructs rather than interchangeable biological-age tests. (Sources: Moqri et al., Validation of biomarkers of aging, PMID 38355974; Belsky et al. comparison of 11 measures, PMID 29149257.)
• Several epigenetic, proteomic, metabolomic, immune and organ-specific measures associate with outcomes in defined cohorts. (Sources: Levine et al. DNAm PhenoAge, PMCID PMC5940111; Lu et al. DNAm GrimAge, PMID 30669119.)
• Functional measures have reproducible prognostic value while remaining nonspecific to aging mechanisms. (Sources: Expert consensus on biomarkers for intervention studies, PMID 39708300; Physical capability and mortality meta-analysis, PMID 20829298.)
• Intended use, population, assay and validation determine what a result can mean. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)

16. What this does not establish

• No comparison identifies a universally “best” biological-age test.
• Age prediction does not prove measurement of aging rate. (Sources: Horvath 2013 multi-tissue clock, PMID 24138928.)
• Outcome association does not prove causality or clinical actionability. (Sources: Levine et al. DNAm PhenoAge, PMCID PMC5940111; Lu et al. DNAm GrimAge, PMID 30669119.)
• Intervention responsiveness does not prove healthspan benefit, lifespan extension or surrogate validity. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)
• This record does not provide purchasing, testing, diagnosis or individual interpretation guidance.

17. Update triggers

Review this record when a consensus framework changes aging-biomarker validation criteria; a major external validation tests calibration across diverse populations; an intervention trial links prespecified marker change to a clinical outcome; a regulatory authority qualifies an aging-related surrogate for a defined use; or a platform materially changes assay comparability. (Sources: FDA: Biomarkers and Surrogate Endpoints; Moqri et al., Validation of biomarkers of aging, PMID 38355974.)