Draft:Biological Age Assessment

Biological age (BA) is an integrated measure used in ageing research to express the extent to which the physiological state of an organism deviates from what would be expected on the basis of its chronological age. While chronological age is closely associated with morbidity and mortality at the population level, at the individual level it fails to capture the substantial physiological, cellular and molecular differences between people of the same age. The mathematical models that quantify biological age are collectively known as aging clocks.

The concept is tied to the central hypothesis of geroscience, according to which the biological process of ageing is the principal integrated risk factor for chronic diseases, including cardiovascular disease, neurodegenerative diseases, type 2 diabetes and cancer. It follows that by targeting the fundamental mechanisms of ageing, the onset of several pathological states could be delayed simultaneously, extending not only lifespan but also the years lived in good health (healthspan).

The methodology of estimating biological age has passed through four distinguishable generations: first-generation clocks trained to predict chronological age were followed by a second generation anchored to mortality risk, then by a third generation measuring the rate of ageing, and finally by a fourth generation of organ-specific, causally disentangled multidimensional omics models.

Conceptual framework

The concept of biological age proceeds from the observation that two people of identical chronological age may differ dramatically in the functional capacity of their organs, the extent of DNA damage, the chronic inflammatory response and the degree of cellular senescence, as a result of individual genetic background, lifestyle and environmental exposures. Biological age is intended to quantify this heterogeneity.

The literature distinguishes three quantities that are frequently conflated:

• Biological age – a state measure expressed in years: the current extent of accumulated physiological damage.
• Age acceleration – the difference between biological and chronological age, or the residual of a regression of biological on chronological age.
• Pace of aging – the rate of ageing, that is, a derivative-like measure of how many biological years elapse per calendar year.

Generations of aging clocks

Generations of aging clocks, their modelling strategy and limitations

Generation	Modelling strategy (anchor)	Main examples	Theoretical and clinical limitations
First	prediction of chronological age	Horvath pan-tissue clock, Hannum clock	the accuracy paradox: the more precisely the clock predicts chronological age, the less information it carries about biological deviation
Second	prediction of phenotypic state and mortality risk	PhenoAge, GrimAge	built on cross-sectional data; measures a state and therefore responds slowly to lifestyle change
Third	pace of aging, longitudinal decline	DunedinPACE	high data requirement: decades-long follow-up studies are needed for training
Fourth	organ-specific and causality-enriched multidimensional omics prediction	Organ Age, ProtAge, causal epigenetic clocks	difficulty of separating damage from adaptation; high technological cost

The PhenoAge anchor

A milestone in the quantification of biological age was the work published in 2018 by Morgan E. Levine and colleagues, which introduced the concept of phenotypic age (PhenoAge). The developers of first-generation epigenetic clocks – among them Steve Horvath and Gregory Hannum – trained machine learning algorithms to estimate chronological age from DNA methylation patterns with the smallest possible error. This led to a conceptual contradiction: if an algorithm predicts chronological age perfectly, then by definition it excludes the individual variance of biological ageing, that is, it loses precisely the ability to discriminate between people of the same age in terms of disease and mortality risk.

Levine and her group therefore shifted the anchor of model training from chronological age to phenotypic mortality risk.

Development

PhenoAge was constructed in two stages. In the first stage, using data from the third NHANES survey, a proportional hazards penalized regression model isolated, from 42 routinely measured clinical biomarkers, the nine that – together with chronological age – most strongly predicted all-cause mortality during the observation period.

The nine PhenoAge biomarkers and their physiological significance

Biomarker	Physiological significance
Albumin (g/L)	marker of hepatic synthetic function; as a negative acute-phase protein its level falls in response to chronic inflammation
Creatinine (μmol/L)	indirect marker of the glomerular filtration rate and thus of kidney function; also related to the age-related loss of muscle mass (sarcopenia)
Fasting glucose (mmol/L)	fundamental indicator of metabolic syndrome, insulin resistance and mitochondrial dysfunction
C-reactive protein (ln[CRP], mg/dL)	acute-phase protein; the principal marker of "inflammaging", the chronic low-grade systemic inflammation of old age
Lymphocyte percentage (%)	marker of immunosenescence; reflects exhaustion of adaptive immunity and involution of the thymus
White blood cell count (109/L)	general activation state of the immune system; responds to occult infection and chronic cellular damage
Mean corpuscular volume (MCV, fL)	marker of ageing of the haematopoietic system; associated with clonal haematopoiesis and with vitamin B12 and folate malabsorption
Red blood cell distribution width (RDW, %)	degree of anisocytosis; a strong predictor of oxidative stress and cardiovascular mortality
Alkaline phosphatase (U/L)	indicator of bone metabolism and of the integrity of the hepatobiliary system

The Gompertz model and back-transformation to years

The mathematical core of the model is the calculation of ten-year mortality risk (the mortality score, $M$) using the Gompertz survival model. The Gompertz function was chosen because it faithfully describes the exponential increase of adult human mortality over time:

$M=1-\exp \left(-\exp(xb)\cdot {\frac {\exp(120\gamma )-1}{\gamma }}\right)$

where $xb$ is the weighted linear combination of the nine biomarkers and chronological age plus a constant, $\gamma$ is the empirically determined scale parameter of the Gompertz distribution, and 120 denotes the ten-year prediction horizon expressed in months.

The resulting $M$ is a probability between 0 and 1, which the algorithm then transforms back into an equivalent age. The conceptual content of phenotypic age is thus the answer to the question: "what is the chronological age of a person in an average reference population whose mortality risk exactly equals the risk computed from the biomarker profile of the individual under examination?"

DNAm PhenoAge

In the second stage of the research, the validated phenotypic age was itself used as an anchor for training a new epigenetic clock: DNA methylation data from whole blood in the InCHIANTI and Women's Health Initiative cohorts were regressed not on chronological age but on the computed phenotypic age. The procedure identified 513 CpG sites that together constitute the DNAm PhenoAge biomarker. Although the clock was developed on blood samples, it tracked ageing closely in almost every human tissue examined.

Transcriptional analysis also revealed the underlying biology: in individuals whose DNAm PhenoAge significantly exceeded their chronological age (phenotypic age acceleration, PhenoAgeAccel), increased pro-inflammatory gene activation – particularly along interferon pathways – was measured, together with reduced activity of the transcriptional and translational machinery, of DNA repair systems and of mitochondrial function. DNAm PhenoAge predicted all-cause mortality, cancer risk, Alzheimer's disease and the decline of physical function more accurately than earlier epigenetic measures. Among later mortality-anchored clocks, GrimAge and its second version showed the strongest predictive performance.

Klemera–Doubal aggregation

While PhenoAge changed the anchor of model building, the Klemera–Doubal method (KDM), published in 2006 by the Czech researchers Petr Klemera and Stanislav Doubal, remains the most robust mathematical framework for the statistical aggregation of clinical biomarkers. The method sharply criticized and superseded the previously dominant procedures, above all multiple linear regression (MLR) and principal component analysis (PCA).

Shortcomings of earlier procedures

In the traditional approach, chronological age served as the dependent variable and the biomarkers as independent predictors. This induces systematic regression toward the mean: because the model minimizes the error of estimating chronological age by least squares, the variance of the resulting biological age estimate is always artificially smaller than the true biological variance of the population.

When principal component analysis is applied, researchers tend to treat the first principal component – which explains the largest share of variance in the data – as the "vector of ageing". This is mathematically unfounded: the first principal component merely indicates the direction of greatest dispersion, which does not necessarily coincide with the trajectory of ageing.

As a consequence of these two biases, the regression slope of both MLR- and PCA-based estimates on chronological age remains systematically below 1 – values reported in the literature are around 0.48 for MLR and 0.62 for PCA. This creates the illusion that physiological decline slows with advancing age, which contradicts gerontological observation.

Inverse regression

Klemera and Doubal recognized that biological age is a latent variable that cannot be observed directly. Instead of making chronological age the dependent variable, they applied inverse regression, treating the observable biomarkers as functions of ageing. For each biomarker $j$ a linear regression on chronological age is computed,

$x_{j}=k_{j}\cdot CA+q_{j}+\varepsilon _{j}$

where $k_{j}$ is the slope, $q_{j}$ the intercept and $\varepsilon _{j}$ the residual error with standard deviation $s_{j}$. The procedure then seeks the biological age that minimizes, in the space of the biomarkers, the error-variance-weighted distance between the individual observations and the regression lines:

$BA_{E}={\frac {\sum _{j}(x_{j}-q_{j}){\dfrac {k_{j}}{s_{j}^{2}}}}{\sum _{j}\left({\dfrac {k_{j}}{s_{j}}}\right)^{2}}}$

The method does not stop there. The algorithm also acknowledges the independent information value of chronological age, particularly in the case of noisy biomarkers, and therefore introduces a Bayesian-type correction that weights the estimate with chronological age as a function of the population variance of biological age ($s_{BA}^{2}$). The chronological-age-corrected estimate – referred to in the literature as KDM2 or method 2 – takes the final form:

$BA_{EC}={\frac {\sum _{j}(x_{j}-q_{j}){\dfrac {k_{j}}{s_{j}^{2}}}+{\dfrac {CA}{s_{BA}^{2}}}}{\sum _{j}\left({\dfrac {k_{j}}{s_{j}}}\right)^{2}+{\dfrac {1}{s_{BA}^{2}}}}}$

The most important property of this construction is that it does not amplify measurement error when several noisy biomarkers are aggregated. Extensive cohort studies have shown that the difference between KDM-derived biological age and chronological age is more strongly and more accurately associated with all-cause and cardiovascular mortality than chronological age alone, outperforming the predictive power of MLR- and PCA-based models. An open-source implementation of the method – together with PhenoAge and a measure of homeostatic dysregulation – is available in the BioAge package for R.

The identical-association assumption

Although KDM is the most accurate aggregation procedure on cross-sectional data, a methodological analysis published in 2024 highlighted a fundamental theoretical problem. Every cross-sectional biological age predictor – including KDM – rests on the so-called identical-association assumption. According to this, true biological age can be written as $BA=CA+\Delta$, where $\Delta$ is individual ageing divergence, and the model builders assume that the strength and direction of a biomarker's association with chronological age directly reflects its association with true physiological decline.

The analysis provided a mathematical proof that this assumption is in principle untestable from cross-sectional population data. If the assumption is violated – that is, if the change of a marker over time is merely a compensatory mechanism rather than a cause of functional decline – then the weight assigned to that marker by KDM becomes uninformative, and the prediction carries essentially no more biological signal than if the weights had been generated at random. This realization opened the way towards longitudinal measures.

Pace of aging

The next theoretical shift in geroscience was to model the rate of ageing instead of its state – that is, a derivative-like ratio expressing how many biological years elapse per calendar year, rather than an absolute value expressed in years.

The Dunedin cohort and DunedinPACE

The approach originated from the Dunedin Multidisciplinary Health and Development Study. Researchers followed a New Zealand birth cohort of about a thousand people born in 1972–73, repeatedly measuring the change of organ-system biomarkers – cardiovascular, metabolic, renal, dental and cognitive indicators – in the same individuals at ages 26, 32, 38 and 45. This design captured not the state of ageing but the actual rate of individual biological decline.

The resulting pace measure was translated into a DNA methylation signature first by DunedinPoAm and subsequently by the DunedinPACE algorithm, based on twenty years of change in 19 biomarkers. The output of the model is not an age but a rate: a value of 1.0 means that the individual's biological ageing is synchronous with calendar time, whereas 1.2, for example, indicates a rate accelerated by twenty per cent. The test–retest reliability of this measure is exceptional: the intraclass correlation coefficient is consistently above 0.90, the highest among validated epigenetic clocks.

The CALERIE trial

The superiority of pace-based models became most apparent in the measurement of interventions. In the CALERIE randomized controlled trial, the effects of two years of mild (11–15%) calorie restriction were tested in healthy young and middle-aged adults. On evaluation, conventional epigenetic clocks measuring age in years showed little or no significant change, whereas DunedinPACE clearly demonstrated that caloric restriction reduced the pace of biological ageing by approximately 11%.

This was a critical methodological insight: estimates describing state, that is accumulated damage, have great inertia and may remain insensitive to short-term interventions, whereas algorithms measuring the rate of ageing can detect the effect of lifestyle change or pharmacological treatment in real time. Biological age is therefore not merely a single number: alongside the state, the direction and the speed of change carry independent information.

Questionnaire-based proxies and lifestyle factors

Processing molecular – epigenetic and proteomic – biomarkers is expensive and technology-intensive, so indices computed from questionnaire data frequently serve as proxies in large population studies.

Dietary inflammatory index

Chronic low-grade systemic inflammation ("inflammaging") is one of the primary pathomechanisms of ageing, and is directly modulated by diet quality. To quantify this, Nitin Shivappa and colleagues developed the Dietary Inflammatory Index (DII), derived from the synthesis of 1,943 independent peer-reviewed sources, including cell biology, animal and human clinical studies.

The energy-adjusted version of the index, E-DII, weights the pro- or anti-inflammatory effect of 45 dietary factors – macro- and micronutrients, fatty acid profiles, fibre, flavonoids, polyphenols and spices – according to their effect on the plasma levels of six inflammatory biomarkers (CRP, IL-1β, IL-4, IL-6, IL-10 and TNF-α). In the calculation, the intake of each factor is standardized against a global mean reference intake and the resulting value is multiplied by the inflammatory effect score of that component. A diet rich in trans fats, refined carbohydrates and saturated fats thus yields a high positive E-DII score, while a diet rich in polyphenols and fibre yields a negative value.

The epidemiological relevance of E-DII is supported by several cohort studies: high values are associated in a dose-dependent manner with elevated high-sensitivity CRP, IL-6 and TNF-α levels. An analysis of the NHANES database found that among patients with chronic kidney disease, those in the most pro-inflammatory tertile of diet faced a 33% higher all-cause mortality risk (HR = 1.33) and a 54% higher cardiovascular mortality risk (HR = 1.54) over five years of follow-up compared with the lowest tertile.

Biases of self-report

Self-reported questionnaire methods carry fundamental cognitive and systematic biases. Comparison with objective measurements – doubly labelled water energy intake assessment or accelerometry – shows that respondents are strongly subject to social desirability bias: they tend to overreport healthy behaviours such as fruit and vegetable consumption or physical activity, while chronically underreporting factors that accelerate ageing, including smoking intensity, alcohol consumption and excessive caloric intake.

The consequence is that questionnaire proxies are well suited to ranking individuals within large populations and identifying cohort-level trends, but without objective blood or tissue biomarkers they are not suitable for the precise clinical determination of absolute individual biological age.

Recent directions (2023–2026)

Validation framework

The proliferation of aging clocks created an unmanageable situation, as divergent validation methodologies made the clocks impossible to compare. The Biomarkers of Aging Consortium was formed to address this, and in consensus documents published in Cell in 2023 and in Nature Medicine in 2024 it laid down the requirements for biomarker development and testing. Validation was divided into three pillars: analytical validation (laboratory reproducibility, handling of batch effects), biological validation (connection to fundamental physiological pathways) and clinical validation (predictive power for health outcomes).

The consortium's open-source Biolearn framework compared 39 aging biomarkers on a unified platform using data from more than twenty thousand individuals. The analysis led to a fundamental finding: the accuracy of chronological age prediction is essentially uncorrelated with a biomarker's capacity to predict mortality. While calendar age was estimated most accurately by the Horvath first-generation skin and blood clock, the second version of GrimAge proved strongest in predicting mortality and functional healthspan. This confirmed Levine's earlier proposition that ageing models should use health outcomes as their anchor, with chronological age serving only as a secondary reference point.

Organ-specific ageing

Interpreting biological age as a single organism-level number is increasingly regarded as obsolete, since the human body ages asynchronously. Researchers at Stanford University combined gene expression databases with plasma proteomic profiles: they first identified some 900 organ-specific proteins whose production in a given organ is at least four times that in other organs, then used machine learning to build separate ageing clocks for 11 major organ systems – brain, heart, lung, liver, kidney, intestine, immune system, arteries, adipose and muscle tissue, and pancreas – using data from more than 5,600 adults.

According to the results, in 18.4% of the population over 50 at least one organ ages measurably faster than the body as a whole, and in 1.7% several organs age extremely fast simultaneously. Accelerated ageing of individual organs also predicts the corresponding pathologies: an accelerated "brain age" may signal the development of Alzheimer's disease and other neurodegenerative disorders up to 15 years before clinical symptoms appear.

Proteomic clocks

The ProtAge clock, a gradient-boosting machine learning model based on 204 proteins, was developed on the plasma proteomic data of more than 45,000 participants of the UK Biobank. The gap between proteomic and chronological age predicted the incidence of 18 different chronic diseases, including decompensated liver cirrhosis, kidney failure and diabetes mellitus.

Compared with the "black box" character of epigenetics, the advantage of proteomic models is that the proteins involved have transparent biological functions. Among the most important contributing proteins of ProtAge is CCL11 (eotaxin-1), a known chemokine of neuroinflammation and systemic inflammaging, while expression of the immune receptor KLRK1 (NKG2D) correlates negatively with ageing, probably because of its role in the immune clearance of senescent cells. Building on this foundation, the Longitudinal Proteomic Aging Index (LPAI) uses data from three consecutive time points and measures not only the momentary state but also the temporal dynamics of the ageing trajectory, correlating closely with frailty and multimorbidity.

Separating damage from adaptation

As omics models became more refined, the concept of causality-enriched biomarkers emerged. Research by Kejun Ying, Vadim N. Gladyshev and colleagues pointed out a fundamental blind spot of conventional aging clocks: the physiological origin of the molecular changes they record is twofold. Some of them are direct signs of degenerative cellular damage, while others represent the organism's protective, compensatory adaptation to that damage. Conventional machine learning models conflated these two opposing processes on purely statistical grounds, since both move together with age. Applying the principles of Mendelian randomization, the authors successfully separated the two components.

The therapeutic significance of the distinction is considerable: if an intervention merely lowers the "number" of epigenetic or proteomic age while erasing the organism's adaptive responses, it does not rejuvenate the body but deprives it of its protection.

Criticism and limitations

• Limitations of cross-sectional data. The great majority of aging clocks are built from single-point data collection. Because of survivorship bias in the cohorts, the biological profile of very old subjects distorts the regression models, since their population has already been selected both genetically and in terms of lifestyle.
• Misestimation in the oldest age groups. PhenoAge is a strong mortality anchor, but clinical testing has shown that the models tend to substantially overestimate biological age above the age of 85; this is caused partly by the scarcity of training data and partly by the linearity of the modelling, which cannot follow the physiological plateau of extreme old age.
• Lack of standardization. Neither the Food and Drug Administration nor the European Medicines Agency has so far issued official guidance on accepting aging biomarkers as clinical endpoints. Filling this gap is precisely the aim of the Biomarkers of Aging Consortium, which must demonstrate analytical reproducibility and responsiveness to interventions.
• Commercial applications. Measurement of biological age has also spread as a commercial service, but a substantial proportion of these tests have not undergone the threefold validation described above, and their results cannot be regarded as a clinical diagnosis.

Adoption in Hungary

The concept of biological age has long-standing roots in Hungarian scientific literature, predating the omics turn: the journal Orvosi Hetilap discussed the significance of biological age in assessing the clinical progression of juvenile idiopathic scoliosis as early as 1979.

The PhenoAge model has direct practical relevance for Hungarian health care, because the underlying clinical markers – albumin, CRP, fasting glucose, full blood count with MCV and RDW, and kidney function – are standard components of the publicly funded routine laboratory panels of Hungarian primary care, so the algorithm could be applied retrospectively to existing patient records as a complement to conventional cardiovascular risk scores.

For Hungarian preventive medicine and dietetics, validating the E-DII index in the Hungarian population is a separate task, given the traditionally high saturated fat content and low fish consumption of the Hungarian diet, which produce a strongly pro-inflammatory profile.

External links

• Biomarkers of Aging Consortium – international consortium developing validation standards for aging biomarkers
• Biolearn – open-source framework for the unified computation and comparison of aging biomarkers
• Biomorg – Biological age – Hungarian questionnaire-based estimation tool built on a PhenoAge anchor and Klemera–Doubal-type aggregation, which also computes a pace of aging value

See also

• Ageing
• Gerontology
• Epigenetic clock
• DNA methylation
• Biomarker
• Life expectancy

Category:Ageing Category:Gerontology Category:Biomarkers Category:Epigenetics