Integrative Biomedical Research

1. Introduction

Ageing, when you strip away the sentiment attached to it, is really just accumulated failure — a slow, uneven erosion of the systems that once kept an organism resilient to insult (López-Otín et al., 2023; Lopez-Otin et al., 2013). It is also, inconveniently, the single largest risk factor for the chronic diseases that dominate modern medicine, and yet clinicians are still largely stuck measuring it with a calendar. Chronological ag

e is easy to record and almost impossible to act on. It tells a physician how many years have passed since a patient was born, but it says remarkably little about how that patient’s kidneys, arteries, or immune system are actually holding up (Li et al., 2022; Polidori, 2024; Satapathy et al., 2026). For decades, public-health guidelines, insurance actuaries, and epidemiological risk models have leaned on this single number anyway, not because it was ever a particularly good proxy, but because nothing more precise was available (Polidori, 2024; Satapathy et al., 2026).

The trouble becomes obvious the moment you compare two people born on the same day. One may still be running marathons at seventy; the other may already be managing three chronic conditions. Genetics, behaviour, socioeconomic circumstance, and sheer environmental luck interact in ways that pull these two trajectories apart, sometimes dramatically, and chronological age has no vocabulary for any of it (Horvath & Raj, 2018; Li et al., 2022; Polidori, 2024). Geroscience’s response to this problem has been to propose an alternative construct — biological age, sometimes called physiological or phenotypic age — intended to capture the actual functional integrity of an individual’s cells, tissues, and organ systems, rather than the number of times the earth has gone around the sun (Yamada, 2025). Decoupling biological age from the calendar matters clinically because it turns ageing into something fluid and, at least in principle, something that can be intervened upon before overt disease appears, rather than after (Li et al., 2022). High-throughput omics technologies have given researchers several candidate ways to estimate this hidden variable — transcriptomic clocks, proteomic clocks, metabolomic signatures — but among these, DNA methylation-based estimators have consistently proven the most robust and reproducible (Ibáñez-Cabellos et al., 2026; Satapathy et al., 2026).

It is worth pausing on why methylation, specifically, turned out to be such a useful signal. Epigenetics broadly refers to chemical modifications layered on top of DNA and chromatin that shape gene expression and cellular identity without altering the underlying sequence (Satapathy et al., 2026). DNA methylation is the best-characterized of these marks: DNA methyltransferases attach a methyl group to the 5′ carbon of cytosine bases that precede guanine, producing 5-methylcytosine at so-called CpG dinucleotides (Satapathy et al., 2026; Wang et al., 2022). The functional consequences depend heavily on location — hypermethylation within CpG islands near gene promoters tends to silence transcription by blocking transcription-factor binding and recruiting chromatin remodellers, whereas broad, genome-wide hypomethylation is more typical of transcriptionally active, open chromatin (Satapathy et al., 2026; Wang et al., 2022). Across the human lifespan, these marks do not simply drift randomly; they shift in a fairly predictable direction, a phenomenon researcher have taken to calling “epigenetic drift” — a genome-wide loss of methylation in late-replicating regions paired with more localized gains at bivalent chromatin domains and Polycomb-repressed promoters (Horvath & Raj, 2018; Galow & Peleg, 2022). Because the underlying DNA sequence itself stays essentially fixed while these methylation marks respond to physical inactivity, chronic stress, diet, smoking, and pollutant exposure, the methylome ends up functioning almost like a running ledger of a person’s lifetime exposures — one that can, remarkably, be mathematically decoded (Galow & Peleg, 2022; Satapathy et al., 2026).

That decoding effort is what gave rise to “epigenetic clocks” — regularized regression models, typically elastic net, that assign predictive weights to a defined panel of CpG sites in order to output an estimated age (Horvath & Raj, 2018; Li et al., 2022). These tools have moved through several distinct generations since their introduction in the early 2010s, and the arc of that evolution is itself instructive. First-generation clocks were trained to predict chronological age directly, using blood or multi-tissue DNAm data; Bocklandt et al. (2011) built the earliest saliva-based estimator from 88 CpG sites in monozygotic twins, Hannum et al. (2013) followed with a 71-CpG blood clock, and Horvath’s (2013) 353-CpG pan-tissue model became the field’s benchmark, validated across fifty-one different human tissue and cell types. These models correlated with chronological age astonishingly well — often above r = 0.90 — but that very success turned out to conceal a limitation. Because the algorithms were optimized purely to track calendar time, they preferentially selected the most temporally stable CpGs, which meant they systematically discarded the more variable, environmentally responsive sites that actually carry information about health (Horvath & Raj, 2018; Levine et al., 2018; Yamada, 2025). Levine and colleagues later described this as the “paradox of chronological ageing”: accuracy against the calendar and biological relevance turned out to be, to some extent, competing objectives (Levine et al., 2018; Satapathy et al., 2026).

Second-generation clocks were built explicitly to correct for this. Rather than training on calendar age, Levine et al. (2018) trained DNAm PhenoAge on a composite “phenotypic age” derived from calendar age plus nine clinical markers spanning immune, metabolic, and renal function. Shortly after, Lu et al. (2019) constructed DNAm GrimAge by regressing methylation data against smoking pack-years and a panel of plasma proteins linked to mortality, including cystatin C, GDF-15, and PAI-1. Both models outperform their first-generation predecessors substantially in predicting all-cause mortality, cardiovascular disease, cancer incidence, and physical and cognitive decline (Li et al., 2022; Satapathy et al., 2026; Yamada, 2025). Third-generation clocks pushed the concept further still, aiming not to estimate a static age at all but to capture the instantaneous rate, or “pace,” of biological ageing. Drawing on repeated biomarker measurements from the Dunedin longitudinal cohort, Belsky and colleagues (2020) introduced DunedinPoAm, later refined into DunedinPACE (Belsky et al., 2022) — a model that behaves less like an odometer and more like a speedometer, and one that appears unusually sensitive to short-term lifestyle change.

Despite this rapid conceptual progress, moving epigenetic clocks from population-level research tools into individual-level clinical instruments remains genuinely difficult. Technical noise across array batches can shift an individual’s estimated age by several years on its own (Apsley et al., 2025; Ibáñez-Cabellos et al., 2026); the methylome is tissue-specific, so a blood-based estimate may say little about pathology unfolding in cortical or cartilage tissue (Li et al., 2022; Satapathy et al., 2026; Yamada, 2025); and most reference datasets remain overwhelmingly drawn from European-ancestry cohorts, raising real concerns about equity when these tools are applied elsewhere (Apsley et al., 2025; Satapathy et al., 2026). What makes the field genuinely exciting, though — and this is really the premise of the present review — is the growing recognition that biological ageing is not a fixed, thermodynamically inevitable process but a plastic, biochemically reversible one (Satapathy et al., 2026). Randomized trials of multidomain lifestyle interventions (Fitzgerald et al., 2021; Olaso-Gonzalez et al., 2026), dietary and physical-activity programs (Fiorito et al., 2021), and even transient cellular reprogramming (Galow & Peleg, 2022) have each demonstrated measurable deceleration, and occasionally apparent reversal, of methylation age. That shift — from clocks as passive descriptors to clocks as therapeutic endpoints — motivates a newer conceptual framework, EpiAge-R (Epigenetic Age with Resilience), which attempts to separate degenerative methylation “noise” from adaptive, repair-associated signal by integrating chromatin topology, long-read nanopore methylation mapping, and mitochondrial dynamics into a single resilience-oriented model (Yamada, 2026).

Guided by this framework, the present review asks three linked questions that we believe should structure the next generation of longevity trials: first, a surrogacy question — can reversal of epigenetic age acceleration serve as a valid surrogate endpoint for hard clinical outcomes such as stroke or cardiovascular events; second, a synchronicity question — does localized tissue rejuvenation translate into systemic biological age deceleration; and third, a resilience question — can a composite resilience index outperform conventional clinical risk scores in predicting recovery from acute physiological stress. In pursuit of these questions, this review sets out four objectives: to compare the accuracy and clinical interpretability of first-, second-, and third-generation clocks across tissues and populations; to synthesize trial evidence on the mechanisms by which lifestyle and pharmacological interventions reverse DNAm age; to outline the EpiAge-R framework as a candidate next-generation vital sign; and, finally, to identify the methodological and ethical gaps that must be closed before biological age can be used responsibly at the level of the individual patient.

2. The Ageing Clock, Measured and Modified: From Epigenetic Biomarkers to Reversible Biological Age

The rapid ageing of the global population is, in a fairly blunt sense, forcing healthcare systems to reconsider what they are actually trying to prevent. Reactive, disease-by-disease management is expensive and arrives late; proactive geroscience — intervening on the ageing process itself — is the alternative on offer, but only if we can measure biological age reliably enough to know whether an intervention is working (Polidori, 2024). Epigenetic clocks, built from DNA methylation, are currently the field’s best answer to that measurement problem (Moqri et al., 2023), and this section works through what is known about them: how they evolved, how they connect to conventional geriatric assessment, how they have been simplified for wider use, and what happens — empirically, not just theoretically — when researchers try to slow them down.

2.1. The Geroscience Imperative

Life expectancy has climbed remarkably over the past century, but healthspan — the years spent free of major chronic disease — has not kept pace, leaving a widening gap filled by late-life frailty, multimorbidity, and polypharmacy (World Health Organization, 2015; Guerville et al., 2020). Traditional medicine has tended to treat conditions like type 2 diabetes, cardiovascular disease, cancer, and neurodegeneration as separate, organ-specific problems (Polidori, 2024). Geroscience pushes back on that framing, proposing instead that ageing itself is the shared upstream driver, and that targeting it could plausibly delay several of these diseases at once (Kennedy et al., 2014). Testing that hypothesis in humans, however, requires biomarkers sensitive enough to register change long before clinical endpoints accumulate — which is precisely the gap epigenetic clocks were built to fill (Moqri et al., 2023).

2.2. Chronological versus Biological Age: The Hallmark Spectrum

Chronological age is simply elapsed time; biological age is meant to capture something closer to true cellular and physiological integrity (Satapathy et al., 2026). Because genetic background, environmental exposure, and lifestyle choices accumulate damage at markedly different rates across individuals, the two measures routinely diverge (Polidori, 2024). At the molecular level, this divergence is underpinned by the now-familiar hallmarks of ageing — genomic instability, telomere attrition, cellular senescence, mitochondrial dysfunction, and epigenetic alteration among them (López-Otín et al., 2023). Epigenetic change occupies a somewhat privileged position among these hallmarks, since the methylome sits at the interface between environmental exposure and the genome, shifting predictably at CpG islands near gene promoters in response to that exposure (Satapathy et al., 2026; Wang et al., 2022). Because it accumulates a record of cumulative stress in this way, the methylome turns out to be an unusually convenient substrate for mathematical modelling of ageing trajectories (Li et al., 2022).

2.3. From Odometers to Speedometers: The Generational Arc of Epigenetic Clocks

Over roughly a decade, DNAm-based age estimators have moved through several conceptually distinct generations, driven largely by advances in penalized regression and machine learning (Apsley et al., 2025; Conole et al., 2025) (Table 1). The earliest models, built by Horvath (2013) and Hannum et al. (2013), were trained directly against chronological age and achieved striking accuracy — yet, as noted above, this came at the cost of filtering out the environmentally responsive CpGs that matter most for health prediction, producing the so-called chronological paradox (Levine et al., 2018). Second-generation models corrected course by training against morbidity and mortality outcomes instead: DNAm PhenoAge combines methylation data with nine clinical biomarkers of physiological dysregulation (Levine et al., 2018), while DNAm GrimAge predicts survival via DNAm-based surrogates for smoking exposure and a set of ageing-associated plasma proteins (Lu et al., 2019; McCrory et al., 2021). Third-generation metrics, epitomized by DunedinPACE, abandon the single-timepoint framing altogether in favour of estimating the instantaneous pace of biological decline from one blood draw, calibrated against nearly two decades of repeated biomarker measurement in a birth cohort (Belsky et al., 2022). Most recently, principal-component (PC)-based recalibrations of these clocks have been introduced specifically to suppress technical noise; by projecting existing models through PCA, Higgins-Chen et al. (2022) reduced within-sample variance between replicate measurements to under a year and pushed intraclass reliability above 0.96 — a meaningful step toward clinical-grade precision (Ibáñez-Cabellos et al., 2026) (Table 1).

2.4. Bridging Scales: From Molecular Signatures to Geriatric Phenotypes

One persistent tension in this literature is the gap between what happens at the level of a single CpG site and what a geriatrician actually observes at the bedside (Diebel & Rockwood, 2021). Clinical frailty assessment tools — the Comprehensive Geriatric Assessment, deficit-accumulation indices, and the Multidimensional Prognostic Index (MPI) among them — quantify functional vulnerability through an entirely different lens than a methylation array does (Diebel & Rockwood, 2021; Polidori, 2024). Encouragingly, the two approaches appear to converge: peripheral-blood expression of AMPKγ1, for instance, declines with age and inversely tracks MPI-defined frailty independent of chronological age (Polidori, 2024). This kind of cross-validation matters, because it suggests molecular clocks and clinical frailty tools are not measuring unrelated phenomena but complementary facets of the same underlying process — and that trials testing anti-ageing interventions could reasonably use both in tandem (Diebel & Rockwood, 2021) (Table 2).

2.5. Democratizing Measurement: Minimalist and Targeted CpG Models

Table 1. Technical and Analytical Specifications of Multi-CpG Epigenetic Clocks. This table catalogues the principal first-, second-, and third-generation epigenetic clocks, including Horvath's, Hannum's, Bocklandt's, GrimAge, and DunedinPACE, among others. For each clock, it lists the source tissue, number of CpG sites used, statistical modelling method, and predicted outcome (e.g., chronological age, phenotypic age, or pace of ageing). It also summarizes each model's key strength and key limitation, drawing on comparative evaluations from Apsley et al. (2025), Ibáñez-Cabellos et al. (2026), Liang et al. (2024), Satapathy et al. (2026), and Yamada (2025), to help readers judge which clock suits a given research or clinical application.

Table 2. Clinical Associations, Pathological Risks, and Biomolecular Mechanisms of Epigenetic Age Acceleration. This table maps how epigenetic age acceleration (EAA) relates to outcomes across ten major clinical domains, including all-cause mortality, cardiovascular disease, cancer, neurodegeneration, metabolic syndrome, pulmonary decline, physical performance, frailty, and Down syndrome. For each domain, it specifies which clock(s) were applied, the reported statistical association (e.g., hazard ratios), and the underlying biomolecular mechanism proposed to explain the link. The synthesis draws on Diebel & Rockwood (2021), Ibáñez-Cabellos et al. (2026), Li et al. (2022), Satapathy et al. (2026), and Yamada (2025), illustrating the breadth of disease processes connected to accelerated epigenetic ageing.Clinical Associations, Pathological Risks, and Biomolecular Mechanisms of Epigenetic Age Acceleration. Summarizes how EAA correlates with clinical outcomes across major disease domains (Diebel & Rockwood, 2021; Ibáñez-Cabellos et al., 2026; Li et al., 2022; Satapathy et al., 2026; Yamada, 2025).

Genome-wide arrays such as those underlying GrimAge or DunedinPACE remain expensive, slow, and bioinformatically demanding — properties that limit their use in large epidemiological studies or routine clinics (Apsley et al., 2025; Ibáñez-Cabellos et al., 2026). In response, several groups have built stripped-down “minimalist” clocks that target only a handful of highly informative CpG sites via inexpensive, high-throughput assays such as pyrosequencing or EpiTYPER mass spectrometry (Gensous et al., 2022; Kim et al., 2025) (Table 3). Weidner et al.’s (2014) three-CpG estimator, built on ITGA2B, ASPA, and PDE4C, achieved a mean absolute error under five years despite its simplicity. Gensous et al. (2022) later validated a six-region panel spanning ELOVL2, NHLRC1, AIM2, EDARADD, SIRT7, and TFAP2E, shown to track accelerated ageing in both Down syndrome and extreme-longevity cohorts. Among these loci, ELOVL2 stands out consistently as perhaps the single most robust methylation marker of age in the human genome, now anchoring several stand-alone clinical and forensic assays (Ibáñez-Cabellos et al., 2026). Kim et al.’s (2025) EpiClock and the sex-stratified ELOVL2 model of Ibáñez-Cabellos et al. (2025) extend this logic further, offering reproducible, low-cost tools suited to population screening even where dense array infrastructure is unavailable (Table 3).

2.6. Reversing the Clock: What Interventional Trials Actually Show

Perhaps the claim most worth dwelling on here is that biological ageing, unlike the arrow of chronological time, appears to be at least partially reversible (Li et al., 2022). Several trials have now put this to the test directly (Table 4). Fitzgerald et al. (2021), in a small pilot randomized controlled trial, exposed healthy adult men to an eight-week multidomain program — an optimized diet, moderate exercise, sleep hygiene, relaxation practice, and probiotic/phytonutrient supplementation — and observed a statistically significant 3.23-year reduction in Horvath DNAmAge relative to controls. In postmenopausal women, the two-year DAMA trial found that dietary intervention slowed DNAm GrimAge, while the physical-activity arm instead reduced the accumulation of stochastic epigenetic mutations within cancer-associated pathways — two distinct mechanisms converging on the same broad outcome (Fiorito et al., 2021). Olaso-Gonzalez et al. (2026), working with frail older adults in Spain, reported that a six-month multidomain program of personalized exercise and nutritional supplementation not only reversed clinical frailty but also reduced DNAm PhenoAge and preserved methylation-estimated telomere length relative to habitual care. At a more mechanistic level, transient expression of the Yamanaka reprogramming factors has been shown to reset the epigenetic age of adult somatic cells almost entirely, reversing markers of cellular senescence without erasing cell identity (Galow & Peleg, 2022) — a striking, if still largely preclinical, demonstration that the ceiling on reversibility may be higher than clinical trials alone would suggest.

2.7. Clinical and Ethical Hurdles

None of this translates smoothly into everyday clinical use just yet. Assay-to-assay technical noise alone can shift an individual’s estimated biological age by several years, undermining single-timepoint diagnostics (Apsley et al., 2025; Ibáñez-Cabellos et al., 2026). Tissue specificity compounds the problem: a blood-derived estimate may simply not reflect what is happening in post-mortem cortical tissue in Alzheimer’s disease, or in localized cartilage in osteoarthritis, which is really just another way of saying that “accelerated ageing” is not a single, synchronized, body-wide process but an asynchronous, organ-specific one (Li et al., 2022; Satapathy et al., 2026; Yamada, 2025). And because the majority of training cohorts remain drawn from European-ancestry, relatively affluent populations, applying these scores uncritically to more diverse groups risks both statistical mismatch and the reinforcement of existing health disparities — particularly if commercial actors such as insurers begin using unvalidated scores for individual risk stratification (Apsley et al., 2025).

2.8. Synthesis and Emerging Directions

Taken together, this body of work reframes biological age from a static, mostly descriptive number into something closer to a modifiable clinical target (Yamada, 2025). The next conceptual step, embodied in frameworks like EpiAge-R, is to move beyond simply registering chronological deficits and instead quantify a system’s capacity to recover from stress — integrating chromatin topology, long-read methylation mapping, and mitochondrial signal into a genuinely multidimensional resilience index (Yamada, 2026). Pairing these molecular tools with established geriatric instruments such as the MPI offers one plausible route toward a validated, reproducible framework for testing longevity interventions

Table 3. Multi-Locus Versus Minimalist and Targeted Epigenetic Clocks: Performance and Validation Parameters. This table contrasts dense, genome-wide array-based clocks (e.g., Horvath's 353-CpG model, Hannum's 71-CpG model) with minimalist, targeted assays that use as few as 1–8 CpG sites (e.g., the Weidner Estimator, Gensous Targeted Clock, EpiClock, and Ibáñez-Cabellos Minimal Model). For each model, it reports the publication year, number of CpGs, assay platform, key genomic loci, and reported accuracy (e.g., MAE or correlation coefficient). The comparison highlights the trade-off between comprehensive genome-wide coverage and the cost-effective, high-throughput potential of low-CpG-count clocks for population screening.

Table 4. Interventions Modulating Epigenetic Ageing, Cellular Hallmarks, and Rates of Biological Decline. This table summarizes interventional evidence, spanning human trials and animal models, evaluating how lifestyle, dietary, pharmacological, and caloric-restriction interventions affect epigenetic age. For each intervention, it lists the protocol duration, study cohort, the specific epigenetic clock used to measure outcomes, and the reported effect size (e.g., years of biological age reduction, with statistical significance where available). Interventions range from multidomain lifestyle programs and vitamin D3 supplementation in humans to caloric restriction and rapamycin treatment in animal models, drawing on Carreras-Gallo et al. (2025), Fiorito et al. (2021), Fitzgerald et al. (2021), Li et al. (2022), Olaso-Gonzalez et al. (2026), and Petkovich et al. (2017).

(Polidori, 2024) — though, as the following sections make clear, getting there will require considerably more methodological standardization than the field currently has.

3. Methods

3.1. Review Design

This article is a narrative review, not a systematic review or meta-analysis — a distinction worth stating plainly at the outset, since narrative and systematic reviews serve different purposes and are held to different reporting standards. Even so, we tried to make the process reproducible wherever a narrative format allows, borrowing structural elements from PRISMA-style reporting (search strategy documentation, explicit eligibility criteria, a traceable screening pathway) so that another reviewer, working from the same protocol, could arrive at a broadly similar evidence base (Moqri et al., 2023). Because the field spans molecular biology, clinical epidemiology, and intervention trials that do not sit neatly under a single controlled vocabulary, a fully systematic approach risked either excluding conceptually important mechanistic and framework-defining papers or drowning the review in low-relevance hits; a structured narrative approach was judged the better fit for synthesizing this heterogeneous literature.

3.2. Information Sources and Search Strategy

We searched PubMed/MEDLINE as the primary database, supplemented by Scopus and Web of Science to capture journals not fully indexed in PubMed, together with Google Scholar for citation-chasing of highly cited papers. The search window spanned January 2011 (the year of the first published epigenetic age estimator) through mid-2026. Search strings combined controlled vocabulary and free-text terms across three conceptual blocks, joined with AND:

• Block A (methylation/clock terminology): “epigenetic clock” OR “DNA methylation age” OR “DNAm age” OR “biological age” OR “epigenetic age acceleration” OR “pace of aging”

• Block B (generational/model terms): “Horvath clock” OR “Hannum clock” OR “PhenoAge” OR “GrimAge” OR “DunedinPACE” OR “DunedinPoAm” OR “CpG clock” OR “elastic net” OR “principal component clock”

• Block C (clinical/interventional terms): “geroscience” OR “longevity intervention” OR “caloric restriction” OR “lifestyle intervention” OR “frailty” OR “mortality” OR “randomized controlled trial”

The full PubMed query took the form: ((“epigenetic clock”[tiab] OR “DNA methylation age”[tiab] OR “biological age”[tiab]) AND (“Horvath”[tiab] OR “Hannum”[tiab] OR “PhenoAge”[tiab] OR “GrimAge”[tiab] OR “DunedinPACE”[tiab] OR “principal component clock”[tiab]) AND (geroscience[tiab] OR “randomized controlled trial”[pt] OR “lifestyle intervention”[tiab] OR frailty[tiab])) AND (“2011/01/01”[dp] : “2026/06/30”[dp]), restricted to English-language records. Reference lists of retrieved reviews and landmark trials were additionally hand-searched to identify papers not captured by the electronic strategy (backward citation searching), and forward citation tracking was performed on foundational papers (Horvath, 2013; Hannum et al., 2013; Levine et al., 2018; Lu et al., 2019) via Google Scholar to identify subsequent validation and application studies.

3.3. Eligibility Criteria

Records were eligible for inclusion if they (a) described the development, validation, or biological interpretation of a DNA methylation-based age estimator in humans or a closely relevant mammalian model; (b) reported clinical, epidemiological, or mechanistic associations between epigenetic age acceleration and a disease, mortality, or frailty outcome; or (c) reported an interventional trial (randomized, pilot, or single-arm) evaluating the effect of a lifestyle, nutritional, or pharmacological exposure on a DNAm clock outcome. Conceptual and framework papers proposing next-generation resilience-based models were also eligible. We excluded records limited to non-methylation biomarkers of ageing (e.g., telomere length alone, transcriptomic clocks alone) unless methylation data were directly compared, conference abstracts without peer-reviewed full text, and non-English-language articles for which a full translation could not be obtained.

3.4. Study Selection and Data Extraction

Titles and abstracts identified through the search strategy were screened for relevance against the eligibility criteria above; full texts of potentially eligible records were then retrieved and assessed in full. For each included study, we extracted, where reported: the clock model and its generation, tissue or biofluid source, number of CpG sites, statistical training method, primary outcome predicted, sample size and cohort characteristics, effect estimates (hazard ratios, correlation coefficients, mean absolute error), and, for interventional studies, the intervention protocol, duration, comparator, and reported change in biological age. These extracted parameters underpin the comparative synthesis presented in Tables 1 through 4 and the effect-size visualization in Figure 2.

3.5. Synthesis Approach

Given the heterogeneity of outcome metrics across clock models (years of age acceleration, hazard ratios, correlation coefficients, mean absolute error), a quantitative meta-analysis was not attempted; instead, findings were synthesized narratively and organized by clock generation (first, second, third, and next-generation/resilience models), consistent with the conceptual framework introduced in the Literature Review. Where multiple studies reported comparable interventional effect sizes (change in biological age in years), these were tabulated together (Table 4) and visualized (Figure 2) to allow qualitative, cross-study comparison, while explicitly avoiding pooled statistical estimates that the underlying heterogeneity would not support.

3.6. Reproducibility Statement

The complete search strings, database platforms, date restrictions, and eligibility criteria are reported above in sufficient detail that an independent reviewer should be able to reconstruct a substantially overlapping literature set; a formal PRISMA flow diagram was not generated because this is a narrative rather than a systematic review, but the screening logic followed the same underlying identification–screening–eligibility–inclusion sequence.

4. Results

Pulling together the molecular, cellular, and clinical-trial evidence surveyed above, a fairly consistent picture emerges: DNA methylation does not drift randomly across the human lifespan but changes in a direction that is, to a striking degree, predictable (Horvath & Raj, 2018; Satapathy et al., 2026). Rather than simple passive decay, these shifts behave more like a cumulative record — a “genomic memory,” if the metaphor is forgiven — of metabolic stress and environmental exposure (Liang et al., 2024; Lu et al., 2019). What follows summarizes what that record reveals once it is decoded across clock generations, clinical outcome domains, simplified assay formats, and, finally, interventional trials.

4.1. The Mathematical and Technical Evolution of Epigenetic Clocks

The generational progression of epigenetic clocks (Figure 1) traces a shift from purely chronological estimators toward increasingly prognostic, and eventually resilience-oriented, indices (Yamada, 2026). First-generation models — Horvath’s 353-CpG pan-tissue clock and Hannum’s 71-CpG blood clock chief among them — used elastic-net penalized regression to map genome-wide methylation directly onto calendar age, achieving correlations generally at or above r = 0.91 (Hannum et al., 2013; Horvath, 2013). As discussed earlier, this accuracy came with a cost: by optimizing strictly for chronological fit, these models tended to discard the more dynamic, health-responsive CpGs that carry information about actual physiological state (Levine et al., 2018; Satapathy et al., 2026).

Second-generation clocks addressed this directly by folding clinical biochemistry and mortality outcomes into the training objective itself (Yamada, 2025). DNAm PhenoAge (513 CpGs) was trained to predict a composite phenotypic age built from ten clinical markers, including CRP, albumin, creatinine, and alkaline phosphatase, using NHANES III data (Levine et al., 2018; Lu et al., 2019). DNAm GrimAge (1,030 CpGs) instead predicted time-to-death via DNAm-based surrogates for seven mortality-linked plasma proteins — including ADM, GDF-15, and PAI-1 — combined with smoking pack-years (Lu et al., 2019; McCrory et al., 2021). Third-generation models pushed further still: DunedinPACE (173 CpGs) abandoned single-timepoint estimation altogether, instead tracking eighteen clinical biomarkers across roughly two decades of longitudinal follow-up in a birth cohort to yield an instantaneous “pace of ageing” readout (Belsky et al., 2022; Liang et al., 2024). Most recently, principal-component recalibration of these existing models has meaningfully improved their technical reliability, compressing within-sample replicate variance to under a year and raising intraclass correlation above 0.96 (Higgins-Chen et al., 2022; Ibáñez-Cabellos et al., 2026). The comparative CpG count, training method, and validated strengths and limitations of each of these models are summarized in Table 1.

4.2. Clinical Phenotypes and Mortality Predictability of Epigenetic Age Acceleration

The clinical relevance of these clocks rests largely on a

Figure 1. Evolution of Epigenetic Clock Generations. These schematic traces the conceptual progression of DNA methylation-based ageing biomarkers over time. It moves from first-generation "odometer" clocks that estimate chronological age, through second-generation phenotypic and mortality-predictive models, to third-generation "pace-of-ageing" clocks that capture the speed of biological decline. The diagram concludes with the emerging resilience-based EpiAge-R framework, illustrating how the field's focus has shifted from simply dating cells to characterizing the dynamics and reversibility of aging.

Figure 2. Change in DNA Methylation-Based Biological Age Across Interventional Trials. This chart compares the reported change in epigenetic age across four representative intervention studies (Fitzgerald et al., 2021; Fiorito et al., 2021; Olaso-Gonzalez et al., 2026; Carreras-Gallo et al., 2025). It includes the habitual-care control arm from Olaso-Gonzalez et al. (2026) as a reference point for natural age progression. The visualization allows direct, cross-study comparison of intervention effects on biological age, despite differences in clock models, cohorts, and protocols used across the trials.

single derived quantity —epigenetic age acceleration (EAA), the residual difference between predicted epigenetic age and actual chronological age (Li et al., 2022; Satapathy et al., 2026). Across large epidemiological meta-analyses, a positive EAA behaves as an independent, dose-responsive predictor of multimorbidity and mortality; each five-year increase in Hannum- or Horvath-derived EAA in blood, for instance, corresponds to roughly an 8–15% increase in the hazard ratio for all-cause mortality (Fransquet et al., 2019). Second-generation clocks outperform first-generation ones substantially in this respect — GrimAge in particular has repeatedly emerged as the single strongest DNAm predictor of remaining lifespan (Lu et al., 2019; McCrory et al., 2021; Satapathy et al., 2026).

The clinical domains in which EAA shows the most consistent signal are summarized in Table 2 and can be grouped roughly as follows:

• All-cause and cardiovascular mortality. GrimAge and PhenoAge acceleration each track independently with mortality hazard, driven partly by DNAm surrogates for inflammatory and vascular-stress proteins such as PAI-1 and leptin (Lu et al., 2019; Yamada, 2025).

• Oncological pathology. Mitotic clocks, which track hypermethylation at Polycomb Repressive Complex 2 (PRC2) target promoters, are sensitive to cumulative cell-division history and correlate with time-to-cancer-onset and cancer-specific mortality (Levine et al., 2018; Li et al., 2022; Yamada, 2025).

• Neurological trajectories. EAA associates with Alzheimer’s disease pathology (amyloid burden, neuritic plaque density), cognitive decline, and neurodegenerative severity (Li et al., 2022; Yamada, 2025).

• Systemic frailty. EAA is consistently elevated among frail individuals as assessed through the Comprehensive Geriatric Assessment, reflecting molecular-level deterioration that parallels — and, in some cohorts, precedes — clinically apparent functional decline (Diebel & Rockwood, 2021; Polidori, 2024).

These associations line up reasonably well with known molecular hallmarks of ageing: transcriptomic profiling of individuals with accelerated phenotypic age tends to show a pro-inflammatory, interferon-skewed signature (“inflammaging”) alongside downregulated DNA-repair, mitochondrial, and translational-fidelity pathways (Galow & Peleg, 2022; Levine et al., 2018; Satapathy et al., 2026).

4.3. Analytical Simplification: Minimalist and Targeted CpG Clocks

Although genome-wide array-based clocks provide comprehensive coverage, their cost and bioinformatic overhead genuinely limit deployment at scale (Gensous et al., 2022; Ibáñez-Cabellos et al., 2026), which is what motivated the minimalist clock models summarized in Table 3. These typically evaluate somewhere between three and eight CpG sites using pyrosequencing, mass spectrometry, or digital PCR. Weidner et al.’s (2014) three-CpG estimator (ITGA2B, ASPA, PDE4C) achieved a mean absolute error under five years despite its simplicity. Gensous et al.’s (2022) seven-CpG targeted clock, spanning ELOVL2, NHLRC1, AIM2, EDARADD, SIRT7, and TFAP2E, reached a correlation of roughly 0.89 with chronological age and a mean absolute deviation of four to six years. Kim et al.’s (2025) EpiClock, built for high-throughput iPlex mass spectrometry across eight CpGs, reported an R² near 0.94 and a mean absolute error of about 3.8 years with strong reproducibility. And because ELOVL2 methylation alone correlates with chronological age at roughly r = 0.92, sex-stratified ELOVL2-only models have achieved mean absolute errors near five years while requiring only a single locus — a genuinely scalable option for population-level or forensic screening (Ibáñez-Cabellos et al., 2025, 2026). These minimalist tools sacrifice some of the noise-averaging benefit that comes from genome-wide, PC-recalibrated models, but in exchange they offer something arguably just as important for translation: affordability and speed (Ibáñez-Cabellos et al., 2026).

4.4. Reversibility of Biological Age: Interventional Trial Evidence

If there is a single finding in this literature most likely to reshape clinical practice, it is that epigenetic age is not a one-way ratchet (Fiorito et al., 2021; Galow & Peleg, 2022). Table 4 compiles the interventional evidence in detail, and Figure 2 visualizes the reported magnitude of change across four representative trials, which together span roughly a half-year to two-year decrease in biological age depending on protocol and clock used.

In frail older adults (mean age approximately 80 years), a six-month supervised multidomain program combining exercise three days per week with daily protein-rich supplementation reversed clinical frailty (SHARE-FI score improved by 2.5 points, p < 0.0001) and produced measurable gains in handgrip strength (p = 0.0053) and balance (p = 0.0031); at the molecular level, this corresponded to a 0.9-year decrease in DNAm PhenoAge in the intervention arm versus a 4.1-year increase among habitual-care controls (p = 0.03), alongside preserved methylation-estimated telomere length (p = 0.0246) (Olaso-Gonzalez et al., 2026). In healthy middle-aged men, an eight-week structured program of diet, exercise, sleep optimization, and probiotic supplementation produced a 3.23-year reduction in Horvath DNAm age relative to controls (p = 0.018), alongside a 15% rise in serum 5-methyltetrahydrofolate and a 25% drop in triglycerides (Fitzgerald et al., 2021). The two-year DAMA trial in postmenopausal women found that dietary intervention alone slowed DNAm GrimAge by 0.41 years (p = 0.01), largely via reductions in DNAm surrogates for PAI-1, leptin, and GDF-15, while the physical-activity arm instead reduced total stochastic epigenetic mutation load by 0.23 mutations (p < 0.0001) within cancer-relevant pathways — evidence that diet and exercise may act through at least partially distinct molecular routes even when both ultimately slow the clock (Fiorito et al., 2021). Sixteen weeks of high-dose vitamin D3 supplementation in vitamin D-insufficient adults produced a further 1.85-year deceleration of Horvath’s pan-tissue clock alongside improved bone mineral density (Carreras-Gallo et al., 2025).

Preclinical work reinforces this picture from another angle. Long-term caloric restriction in rhesus monkeys and mice delays methylation drift and preserves chromatin accessibility and sirtuin pathway activity (Li et al., 2022; Petkovich et al., 2017), and chronic pharmacological modulation of nutrient-sensing pathways — rapamycin-mediated mTOR inhibition, for instance — produces comparable epigenetic deceleration in animal models (Li et al., 2022; Petkovich et al., 2017). Taken together, the human and preclinical trial evidence summarized here supports treating epigenetic clocks not merely as descriptive biomarkers but as genuinely dynamic, intervention-sensitive endpoints — a foundation that the Discussion section builds on directly.

5. Discussion

5.1. From Passive Odometer to Active Intervention Target

The central argument of this review, if it has to be compressed into one sentence, is this: epigenetic clocks have quietly shifted from being a way of describing ageing to being a way of testing whether we can actually do something about it. The generational progression documented here (Figure 1) — chronological, then phenotypic, then pace-of-ageing, and now resilience-oriented — is not merely a story of improving statistical fit. Each transition reflects a deliberate change in what the model is being asked to predict, moving steadily closer to outcomes that matter to patients rather than outcomes that are merely easy to fit (Levine et al., 2018; Yamada, 2025). That the interventional trials summarized in Table 4 and visualized in Figure 2 can now detect deceleration on the order of months to a few years within trial durations of weeks to months is, frankly, a level of sensitivity that would have seemed implausible for a biomarker a decade ago (Fitzgerald et al., 2021; Olaso-Gonzalez et al., 2026).

5.2. Reconciling Divergent Mechanisms Across Trials

One pattern worth dwelling on is that different interventions do not appear to move the clock through the same molecular door. Dietary optimization in the DAMA trial slowed GrimAge specifically through reductions in inflammatory and metabolic protein surrogates, whereas physical activity in the same cohort instead reduced stochastic epigenetic mutation load in largely separate pathways (Fiorito et al., 2021). If this distinction holds up in larger trials, it has a fairly direct clinical implication: multidomain interventions may outperform single-component ones not simply because they do “more,” but because diet and exercise appear to be correcting genuinely different aspects of epigenetic dysregulation — one stabilizing systemic inflammaging, the other reinforcing DNA-repair and antioxidant capacity (Fiorito et al., 2021; Wang et al., 2022). The multidomain trials in Table 4 (Fitzgerald et al., 2021; Olaso-Gonzalez et al., 2026), which combine several of these components simultaneously, do report some of the larger effect sizes in the dataset, which is at least consistent with this reasoning, though a formal head-to-head comparison of single- versus multi-component protocols has not yet been conducted.

5.3. The Persistent Problem of Tissue and Population Specificity

It would be dishonest to present this evidence without also flagging its limits squarely. Most of the clock models synthesized in Table 1 were built and validated in blood, yet ageing is not a synchronized, body-wide process — a blood-derived clock may simply be blind to what is happening in cortical tissue, cartilage, or other organ-specific compartments (Li et al., 2022; Satapathy et al., 2026; Yamada, 2025). Layered onto this is the reference-population problem raised in the Literature Review: because the majority of training cohorts remain drawn from relatively affluent, European-ancestry populations, applying these clocks to more diverse groups risks both reduced accuracy and, more troublingly, the reinforcement of existing health inequities if unvalidated scores are ever used for individual-level risk stratification by insurers or employers (Apsley et al., 2025). We would argue that this is not a peripheral caveat but a central constraint on how quickly this technology can, or should, move into everyday clinical use.

5.4. Toward Resilience-Based Frameworks: EpiAge-R and Beyond

The EpiAge-R framework represents, we think, a genuinely useful reframing of the field’s central question. Rather than asking only “how much biological damage has accumulated,” it asks “how well can this system recover” — integrating chromatin topology, long-read methylation mapping, and mitochondrial signal into a composite resilience index (Yamada, 2026). This distinction matters clinically because damage and recovery capacity are not the same construct; two patients with identical EAA scores could plausibly have very different trajectories depending on their underlying repair capacity, and a purely damage-oriented clock has no way to distinguish them. Whether EpiAge-R, or a similar resilience-oriented model, actually improves on existing clocks in predicting outcomes such as post-surgical recovery remains an open empirical question — one of the three hypothetical questions this review proposed at the outset — but the conceptual shift it represents seems, to us, like the right direction of travel.

5.5. Limitations of the Present Review

This review is narrative rather than systematic, and while we followed a documented, reproducible search protocol (Section 3), we did not conduct formal risk-of-bias assessment or quantitative meta-analysis, given the heterogeneity of outcome metrics across the included studies. Several of the interventional trials summarized in Table 4 (Fitzgerald et al., 2021; Carreras-Gallo et al., 2025) are small, single-arm, or pilot studies, and effect sizes drawn from such trials should be treated as hypothesis-generating rather than confirmatory. Larger, adequately powered, and more demographically diverse randomized trials — ideally pre-registered with epigenetic age acceleration as a defined secondary or surrogate endpoint — will be needed before firm clinical recommendations can follow.

6. Conclusion

The explosion of work on epigenetic clocks over the past decade raises a question that the field is only now positioned to answer: whether biological age, as these tools measure it, is something clinicians can act on rather than merely observe. Taken together, the evidence reviewed here suggests the answer is cautiously yes. Clocks have moved from crude chronological estimators into biologically meaningful predictors of morbidity, mortality and frailty, and — perhaps more consequentially still — into endpoints sensitive enough to register the effects of lifestyle and pharmacological intervention within weeks to months rather than years. Second- and third-generation models consistently outperform their predecessors on clinically relevant outcomes, minimalist targeted-CpG assays are beginning to make biological age screening genuinely affordable outside specialist laboratories, and a growing set of trials demonstrates that this clock can, to a meaningful degree, be slowed and in some cases apparently turned back. That promise should not be overstated. Technical noise between array batches, the tissue specificity of the underlying methylation signal, and the persistent over-representation of European-ancestry, relatively affluent cohorts in training data remain genuine obstacles to using any single clock as an individual-level diagnostic today. Resilience-oriented frameworks such as EpiAge-R, which aim to separate degenerative methylation noise from adaptive repair capacity rather than treating ageing as pure accumulated damage, offer what we think is the right conceptual direction for the field’s next phase — but they, like the trials summarized here, will need considerably larger and more diverse validation before biological age can responsibly inform decisions made about, or by, an individual patient.

Author Contributions

M.B. conceived and designed the review, conducted the literature search, analyzed and interpreted the relevant literature, and drafted the manuscript. R.K.A. contributed to the conceptual development of the review, interpretation of the literature, and critical revision of the manuscript. Both authors reviewed and approved the final version of the manuscript and agreed to be accountable for all aspects of the work.

Acknowledgements

The authors would like to thank the Department of Pharmacy, Kalinga University, Naya Raipur, Chhattisgarh, India, for providing academic and institutional support during the preparation of this manuscript. The authors also acknowledge the contributions of researchers whose published work provided the scientific foundation for this review.

References