Can a blood test measure the age of individual organs?

Organ proteomic clocks have value for research and risk prediction, while tissue origin, age scales and intervention response require separate validation. We audited public model coefficients and original supplementary tables, then built reference-population, multiple-threshold and protein release–clearance models to examine how these readouts arise and how they can be misinterpreted.

01How a number becomes an “organ age”

Oh and colleagues' 2023 study used GTEx tissue RNA data to identify genes expressed at least four times more highly in one organ than in any other. These organ-enriched genes were mapped to plasma protein measurements. Sets assigned to the same organ were then used to train models predicting chronological age. Original study

This introduces biological information, but the training target remains chronological age. The known answer was not a directly measured “true age” of cardiac pumping, renal filtration or brain pathology. The main 2023 models were trained in 1,398 cognitively unimpaired people and tested in other cohorts. The 2025 Olink versions were newly trained in UK Biobank. 2023 original tables 2025 study

In a linear LASSO model, predicted age is an intercept plus a weighted sum of standardized protein measurements; the 2023 versions also include a sex term. The subsequent age gap commonly subtracts the expected prediction for same-aged people and is divided by a relevant standard deviation. A reported age gap therefore is not always predicted age minus chronological age, and one standard deviation does not represent a fixed number of years across models.

02What proteins do the models actually use?

We linked the public 2023 SomaScan and 2025 Olink coefficients to each study's own protein annotations and candidate input panels. Similar organ labels can describe substantially different retained measurements. Coefficient tables Olink coefficients and methods

For example, six protein measurements have nonzero mean coefficients in the 2023 heart ensemble. The 2025 version has only two, with the NT-proBNP measurement mapped to NPPB accounting for approximately 97.8% of total absolute protein coefficient mass. Its kidney model has four measurements, with REN accounting for approximately 61.3%. The age-trained brain models have 201 and 30, respectively. The 2023 counts do not mean that every bootstrap submodel uses the same complete set. That paper also contains a cognition-optimized brain model, which is a different clock.

These percentages describe concentration of standardized coefficients, not percentages of organ aging explained or variance explained. Correlated proteins can share information, while measurement noise and platform-specific standardization also affect coefficients.

NT-proBNP is itself associated with cardiac load and clinical cardiac state. A model with a large weight on it may capture useful cardiac risk information. The label “heart age” alone does not establish that the readout measures all aspects of cardiac function.

Annotation adds another detail: some measurements target protein complexes whose components have different enrichment annotations. We checked those component-level matches without treating a complex's organ label as evidence that every component has one exclusive source.

03Why can an unchanged sample move with its reference population?

The original 2023 analysis calculated relative gaps within cohorts. The public software defaults to training-cohort standardization and reference curves, with an option to recalibrate within the target cohort. The 2025 application tutorial standardizes target-cohort protein data and fits age references using either a population sample or its controls. SomaScan software Olink application tutorial

A fixed reference supports tracking on one scale but requires handling platform and batch differences. Recalibration in a new cohort can remove technical shifts while also removing real population differences. These are different measurement questions, and the reference rule needs to accompany the result.

Using the actual 2025 heart-model weights, we constructed a hypothetical example. Two groups have the same age distribution, but both protein measurements in the second group are 1.5 baseline standard deviations higher. Age loadings and within-group variability are explicitly assumed. The index sample always retains the same protein measurements; only the proportion of the shifted group in the reference population changes.

Under a fixed training reference, its standardized score is approximately 1.61. With whole-target-cohort recalibration, it becomes approximately 1.44, 0.63 and 0.14 when the shifted group constitutes 5%, 50% and 90% of the population. Nothing changes in the index sample.

An easily missed step is final standardization. Even if controls define the raw age gap, subsequently z-scoring all participants recentres the score on the entire sample. In our age-balanced example, the two routes give identical final standardized scores.

These are neither patient results nor a fitted disease model. They show why comparison of personal readouts requires consistent model versions, preprocessing and reference rules: a decline in reported years may partly reflect a change in the measuring scale.

04Looking at multiple organs creates more tail labels

Calling a score above 1.5 or 2 standard deviations “extreme aging” first defines a location in a distribution. The threshold does not validate a clinical diagnosis by itself.

As a simple reference, if 11 scores are independent standard-normal variables, approximately 22% of individuals have at least one score above positive 2 SD. This calculation does not require adding a distinct subgroup with unusually rapid aging in one organ.

We also used the published 2025 age-gap correlation matrix to construct a reference that preserves its correlations while assuming normal marginal distributions. Following that paper's classification scope, this includes 11 organs plus the Organismal score, for 12 scores in total. Each setting uses 300,000 simulated individuals. Correlation matrix and thresholds

At 1.5 SD, approximately 46% of this reference population has at least one positive extreme score, and approximately 78% has at least one positive or negative extreme score. Increasing the threshold or changing the correlation structure changes those proportions.

This is not an estimate of clinical misdiagnosis or evidence that organ aging does not exist. Normal marginals are an assumption; real distributions may be skewed. We also did not reproduce the 2023 paper's cluster-based ageotypes. The calculation shows that label prevalence needs to be interpreted with the number of scores, thresholds and correlations. Health significance requires separate functional and outcome evidence.

05Release and clearance stand between tissue origin and blood concentration

RNA enrichment in an organ is a useful source clue, but it does not trace every circulating molecule to its origin. Proteins can be actively secreted, retained locally, or enter blood through injury, turnover and barrier changes. The human secretome resource itself distinguishes such destinations. Human secretome study

The 2025 study also compared RNA-derived enrichment with tissue proteomics. Approximately 80% of the relevant proteins received concordant classification in the same organ; others were enriched elsewhere or not classified as enriched. Tissue expression, final protein distribution and actual plasma origin are different measurements. Methods and supplementary figures

To illustrate a basic identifiability problem, we used a one-compartment model: concentration changes at the input rate minus the clearance-rate constant times current concentration. With fixed volume and constant parameters, steady concentration equals input rate divided by the clearance-rate constant.

Doubling input and halving the clearance-rate constant both double steady concentration. Doubling both input and clearance can leave concentration unchanged. A single identical concentration cannot distinguish those processes. A known input pulse followed by measurements over time can provide additional kinetic information in this idealized system.

Rates and time units are assumed; no human protein half-life is fitted. This model also does not establish that clearance is the main cause of any real clock's signal.

Empirical evidence likewise requires both sides. Plasma NfL is associated with renal-function measures, but a study examining blood, cerebrospinal fluid or brain imaging found that creatinine adjustment did not materially change NfL's associations with neurodegeneration measures. Clearance is a mechanism to test, not a conclusion established by correlation alone. Kidney function may also relate to actual disease processes. NfL and renal function

06What does validation across populations validate?

A separate study by Wang and colleagues trained LightGBM models in UK Biobank and evaluated them in China's CKB and the US Nurses' Health Study. It provides valuable evidence across populations, using different protein selection, algorithms and calibration from the LASSO clocks above. Cross-population study and original tables

We recalculated the correlation between the published model-performance profiles: approximately 0.975 for UKB versus CKB, 0.959 for UKB versus NHS, and 0.928 for CKB versus NHS. These correlations compare the performances of 11 models. They do not mean that every organ's age prediction has almost 98% individual accuracy.

For individual models, brain-age correlations with chronological age are approximately 0.772, 0.760 and 0.614 in the three populations. Heart-model correlations are approximately 0.362, 0.454 and 0.226. Correlation also depends on age range: identical prediction noise can produce higher correlations in a population with a wider age span.

These external samples are not three equivalent general-population surveys. The CKB sample comes from an ischemic-heart-disease case–cohort design, while the NHS sample comes from nested sampling for a colon-cancer study and includes only women. Cross-population results remain valuable, while age prediction, absolute calibration and transport of clinical risk require separate assessment.

07Disease prediction has support and failures to replicate

Similar names do not make findings completely independent or interchangeable replications.

The 2025 Olink study found strong prognostic associations for brain and immune scores for some outcomes, alongside only moderate-to-strong agreement between platforms. Longitudinal analyses required separately trained models covering fewer proteins. Approximately 68% of those initially labelled extreme no longer met the same extreme threshold at a later measurement. Technical change, genuine biological change and regression to the mean can all contribute; leaving an extreme category is not direct evidence of organ rejuvenation. Olink study

The 2026 EPIC study analysed 17,473 participants using a case–cohort design with design-appropriate weighting. It supported associations of organ scores with certain corresponding diseases and mortality, but did not replicate clear neurodegenerative-outcome prediction with the 2023 SomaScan brain model it evaluated. Platform, model version, entry age, outcomes and follow-up duration may contribute to the difference. These data cannot isolate one explanation. EPIC final publication

Consequently, the brain clock cannot be described as the strongest predictor in every independent study.

08Predictive improvement and clinical benefit remain separate questions

EPIC compared discrimination of mortality models. Recalculation from its actual source table shows that adding the Global proteomic age gap to a model with age, sex, center and lifestyle risk factors raises concordance from approximately 0.730 to 0.739. Adding all organ and Organismal scores yields approximately 0.743. Original figure values and methods

The latter increment is approximately 0.0135. It describes discrimination of outcome ordering, not 1.35% fewer deaths. The figure retains the authors' bootstrap intervals. Since models share participants, overlap of two intervals does not replace a paired test of their difference.

We did not obtain a participant matrix that permits these analyses to be rerun, and did not repeat case–cohort weighting, model fitting or the authors' bootstrap. Public summary tables support arithmetic checks but cannot independently establish calibration, clinical thresholds or decision net benefit. Even statistical improvement leaves the question of whether the model changes appropriate clinical decisions and ultimately improves outcomes.

09Finer cell labels do not automatically resolve these limits

Ding and colleagues' 2026 study extended protein clocks to cell-type-enriched panels, examining disease and mortality across SomaScan, Olink and multiple cohorts. It moves to a finer cellular level while still relying primarily on RNA enrichment to label models. Cell-type study

The authors explicitly acknowledge that some proteins assigned to particular cells may be produced or released more broadly under other conditions. Some outcome associations persist after renal adjustment, so all signals cannot simply be attributed to clearance. At the same time, analyses using existing Knight-ADRC and UKB populations cannot count as wholly new independent population replications.

Just as sequence counts and function require an additional evidential connection in our immune-diversity study, finer protein labels need testable links to actual sources, tissue function and clinical outcomes.

10What can these readouts best answer now?

Organ proteomic clocks can investigate population heterogeneity, generate mechanistic hypotheses and test whether they supplement existing risk models. They have outcome evidence beyond age fitting, while remaining short of a common physiological-age ruler for all organs.

The most informative next studies would combine fixed model versions and reference rules with repeated sampling and organ-function or tissue measurements, and evaluate clinical prediction in independent populations. Intervention research must additionally test whether treatment-induced score changes reliably predict clinical outcomes, including whether a score can change through pathways that do not improve health.

Our original-table audit and conditional models explain how readouts arise and which processes a single blood draw may not separate. They do not recover anyone's true organ age or estimate life extension.

Download code, public numerical inputs, original figures and verification results