Skip to main content
v2026.11,986 entries · CC-BY 4.0

Biomarkers of Aging: Measures and Evidence

Compare aging biomarker inputs, training targets and outputs, then separate laboratory reliability from outcome prediction and surrogate evidence.

Written and maintained by CASRAI Editorial Board

Published

On this page

Biomarkers of aging are measurable biological characteristics used to study age-related change, differences in health risk, or the pace of physiological decline. They include clinical laboratory composites, DNA methylation models and plasma protein signatures. Their meaning depends on what they were trained to measure and which use the evidence supports.

Short answer

An aging biomarker is useful only in relation to a defined research question. Predicting chronological age, predicting later disease or death, and detecting a response to an intervention are different tasks. A model that performs well at one task has not thereby demonstrated the others. The original studies by Horvath, Levine, Lehallier and Belsky illustrate these different approaches.

The comparison dataset below connects five named measures to their inputs, training targets and outputs. It separates clinical phenotypic age from DNAm PhenoAge, and separates an age estimate from a pace measure. For laboratory and study teams, those distinctions come before interpreting a number as evidence that aging has slowed. The checklist later in this guide applies FDA context-of-use principles to the research examples; it is an interpretation aid, rather than an aging-test qualification standard.

Definition and scope

The FDA describes biomarkers as measured characteristics that indicate health, disease, or a response to an exposure or intervention. Aging research applies that broad idea to changes across the lifespan and to differences between people of the same chronological age. In Levine and colleagues’ research, the aim was to capture differences in disease onset, functional decline and mortality that chronological age alone does not describe.

Chronological age is elapsed time. Biological age, when reported by one of these models, is an estimate constructed from a selected set of measurements and a particular training objective. Pace of aging concerns change over time: DunedinPACE was trained against longitudinal changes in organ-system indicators. These outputs answer related questions, but their numerical scales and underlying targets differ.

The broader scientific rationale belongs to geroscience and the study of shared aging mechanisms. Measurement also differs from a mechanistic classification. The hallmarks of aging framework organizes biological processes; the biomarkers described here are measured readouts or statistical combinations of readouts. Calling a process a hallmark does not specify which assay or model should represent it in a study.

Keep the biomarker separate from the outcome it might help predict. FDA distinguishes clinical outcomes that directly concern how people feel, function or survive from surrogate endpoints used in their place. An age-related molecular signal can be valuable research evidence without being an established substitute for those outcomes. This guide concerns study interpretation, rather than individual diagnosis or treatment.

How biomarkers of aging are used in practice

A study first generates measurements, then applies a defined analysis, and finally interprets the result against its research objective. Horvath’s work combined methylation measurements with coefficients selected through elastic-net regression. Lehallier and colleagues measured plasma proteins using an aptamer platform before building age-prediction models. The assay and the model are distinct parts of the evidence chain.

Start with the training target

Horvath’s multi-tissue model learned a relationship between methylation and chronological age. Levine’s clinical phenotypic-age model instead incorporated mortality information when selecting clinical variables, then expressed the resulting composite in years. DNAm PhenoAge was trained in a second stage to predict that clinical phenotype from methylation. DunedinPACE used longitudinal physiological change as its target.

These choices affect the interpretation of validation results. Agreement with chronological age addresses age prediction. Association with subsequent mortality addresses prognostic information. Repeat measurements address technical reliability. Each result adds evidence, but none should be relabeled as another kind of validation. Horvath explicitly notes that correlation with age does not establish calibration: a predictor can track differences between people while its numerical estimates remain systematically displaced.

A trial example: different readouts, different findings

Waziry and colleagues conducted a post hoc methylation analysis using blood samples from the randomized CALERIE trial. They reported a small intervention effect on DunedinPACE, without significant changes in biological-age estimates including PhenoAge and GrimAge. Their principal analysis used principal-component versions of those two clocks, a detail that matters when comparing the finding with analyses using the original algorithms.

The authors also state that conclusive evaluation of healthy-aging benefit requires longer follow-up for outcomes such as chronic disease and mortality. The result therefore illustrates why “a biomarker changed” and “clinical benefit was established” are separate statements. It also shows why a report should identify its exact algorithm instead of treating all biological-age measures as interchangeable endpoints. The trial is an evidence example, not a recommendation to undertake its intervention.

Clinical laboratory composites

Levine and colleagues’ clinical phenotypic age combines nine clinical biomarkers with chronological age. The inputs include albumin, creatinine, glucose, C-reactive protein and several blood-cell measures. The authors developed the score using NHANES III and assessed it in NHANES IV. Its construction links routine biological measurements to an outcome-informed composite, rather than treating any one laboratory value as a complete aging measure.

Clinical phenotypic age and DNAm PhenoAge share a development history but require different inputs. The first uses clinical measurements; the second estimates the phenotype through DNA methylation at 513 CpG sites. A study record that simply says “PhenoAge” can leave that distinction unresolved. Naming the implementation makes the result interpretable and identifies what data were actually collected.

DNA methylation age models

Horvath’s original multi-tissue predictor selected 353 CpG sites from datasets spanning 51 healthy tissues and cell types. Its breadth was part of the study design, not a property shared automatically by every methylation model. The training data, selected sites, preprocessing and fitted coefficients together define the predictor. The same assay output can therefore support different models with different objectives.

Age acceleration describes a departure of estimated age from chronological age, but a report still needs to specify how that departure was calculated. Horvath discusses both age differences and age-adjusted analyses. An estimated age and an adjusted departure are different reported quantities; naming the quantity prevents a reader from confusing an analysis variable with the model’s direct output.

Methylation measures of pace

DunedinPACE is also based on methylation, yet its target differs from a chronological-age clock. Belsky and colleagues used changes in 19 organ-system indicators measured at four time points across two decades, then trained a blood methylation predictor of that longitudinal pace. The final algorithm uses 173 CpG sites and can be applied to a single-timepoint methylation sample.

The distinction is between the history used to train the model and the sample needed to apply it. One blood sample can produce a pace estimate because the model learned from repeated measurements in its development cohort. It does not mean that the laboratory directly observed that participant’s long-term physiological trajectory from one sample. The underlying design is explained further in longitudinal study design.

Plasma protein signatures

Lehallier and colleagues measured 2,925 plasma proteins in 4,263 people aged 18–95. Their study identified substantial nonlinear changes with age and tested a 373-protein age-prediction model. This is a separate measurement family from methylation: it observes circulating proteins using SomaScan aptamer technology.

The nonlinear findings also discourage assuming that every age-associated protein follows a steady, constant trajectory. A cross-sectional association between protein levels and age describes the sampled population. It does not, by itself, directly measure the within-person rate of change that a longitudinal study observes. The dataset below preserves the distinction between this age-prediction approach and DunedinPACE’s longitudinal training target.

Where biomarkers matter for laboratories and study teams

For a laboratory, the relevant evidence extends beyond the name of the biomarker. FDA’s qualification transcript states that preanalytical considerations and test performance must support the intended context of use. Its context-of-use transcript identifies the biomarker, measured aspect, population, purpose and interpretation or action as relevant considerations. The following checklist applies those principles to the aging studies described here.

  • Identify the material and measurement. Record whether the study uses clinical laboratory values, blood DNA methylation or plasma protein measurements. Levine’s clinical and methylation models demonstrate why a shared label does not establish a shared input.
  • Name the exact model and processing approach. Record coefficients or algorithm version and preprocessing. CALERIE’s use of principal-component clocks illustrates why the implementation belongs in the methods and results.
  • Match the population and setting to the evidence. Describe the development and validation samples. Horvath’s multi-tissue work and DunedinPACE’s blood-based development support different statements about the setting in which each model was studied.
  • Keep technical performance visible. Belsky and colleagues excluded low-reliability methylation probes during development. The CALERIE analysis processed repeated samples from a participant together and included both study groups on chips to limit technical confounding.
  • Specify the intended interpretation. State whether the result will describe age, predict an outcome or evaluate change after an intervention. FDA’s framework makes that intended use central to the evidence needed.

These checks help distinguish a biological hypothesis from a measurement decision. They do not establish particular specimen containers, storage conditions, collection intervals or purchasing specifications. Those details would require the assay’s own validated documentation; the cited aging studies are not universal laboratory supply specifications.

Illustrative composite: a study-use statement

An illustrative study record might read: “Measure change in the specified blood methylation pace algorithm between study visits, compare the randomized groups, and retain later clinical outcomes as separate endpoints.” This is an editorial example of documenting purpose, not a validated protocol. Its value is that it identifies the readout, comparison and boundary of interpretation before results are available.

It leaves several evidence questions visible: whether technical variation is sufficiently controlled, whether the model has been evaluated in the study population, and whether a detected change predicts meaningful benefit. Those questions require separate support. For the last step, see surrogate endpoint validation in clinical trials.

Common misunderstandings

“A biological-age number is a direct reading of one underlying true age.” The cited models were trained against different targets. Their labels and units do not make them equivalent measurements. A useful report explains what the number represents before interpreting its magnitude or comparing it with another model.

“High age correlation proves suitability for an intervention trial.” Horvath distinguishes correlation from calibration, while CALERIE demonstrates that different readouts can produce different intervention findings. Age prediction, responsiveness and clinical-benefit prediction each need their own evidence.

“Reliable measurement establishes clinical meaning.” Technical repeatability concerns the measurement process. FDA states that both analytical and clinical validation support qualification, within the specified context. A reproducible output still needs evidence for the interpretation assigned to it.

“Qualification approves the test for patient care.” FDA explicitly distinguishes biomarker qualification in drug development from approval or clearance of a diagnostic device and from use in clinical practice. A paper describing a research measure is not evidence of any of those regulatory determinations.

“A favorable biomarker change proves longer life.” FDA cautions that surrogate endpoints can fail to capture overall benefit and risk, even when useful in a particular setting. Waziry and colleagues likewise call for clinical-outcome follow-up. Report the observed biomarker effect and its supported interpretation without converting it into an unmeasured survival benefit.

Sources

Data: Aging biomarker measures: inputs, training targets and outputs

Original cross-study compilation of five aging measures from four primary development papers. It distinguishes the measured inputs, model targets and reported outputs; the rows describe the cited implementations, not all variants or regulatory qualifications.

Aging biomarker measures: inputs, training targets and outputs
MeasureMeasured inputTraining targetOutputEvidence boundarySource
Horvath multi-tissue clockDNA methylation at 353 CpG sitesTransformed chronological ageDNA methylation ageDeveloped from datasets spanning 51 healthy tissues and cell types; age correlation alone does not establish calibration.Horvath 2013; PMC4015143
Clinical phenotypic ageNine clinical biomarkers plus chronological ageMortality-informed phenotypic-age modelPhenotypic age in yearsDeveloped using NHANES III and validated using NHANES IV; distinct from the methylation predictor.Levine et al. 2018; PMC5940111
DNAm PhenoAgeBlood DNA methylation at 513 CpG sitesClinical phenotypic ageDNA methylation estimate of phenotypic ageA second-stage methylation model; its input is not the nine clinical biomarker measurements.Levine et al. 2018; PMC5940111
Plasma proteomic age modelSomaScan plasma protein measurements; 373-protein predictorChronological agePredicted ageStudy measured 2,925 proteins in 4,263 people aged 18-95; this age model is not a direct longitudinal pace measurement.Lehallier et al. 2019; PMC7062043
DunedinPACEBlood DNA methylation at 173 CpG sitesChange in 19 organ-system indicators across four time points and two decadesPace-of-aging estimateSingle-timepoint methylation application trained on longitudinal change; evaluated in five additional datasets.Belsky et al. 2022; PMC8853656

Follow CASRAI

Research-administration guidance, standards updates and independent tool reviews.

Ask CASRAI · free to try

Ask about Biomarkers of Aging: Measures and Evidence

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →