Biological-Age Clocks

MRI-based brain age is a biological-age estimate derived from structural or functional brain imaging features—including cortical thickness, white-matter integrity, grey-matter volume and functional connectivity—processed by machine-learning models trained on large neuroimaging cohorts. The gap between predicted brain age and chronological age, termed the brain-age gap or BrainAGE (introduced by Franke et al., 2010), is a biomarker of brain health: a positive gap (brain appearing older) associates with cognitive decline, neurodegeneration, stroke and all-cause mortality, while a negative gap is linked to better cognitive reserve. Accuracy and interpretability vary with imaging protocol, preprocessing pipeline and training cohort demographics.

CausAge is an epigenetic-age clock introduced by Ying, Gladyshev and colleagues (preprint 2022; Nature Aging 2024) that attempts to address a fundamental limitation of correlation-trained clocks: standard elastic-net or regression clocks select CpG sites associated with age without distinguishing whether the methylation change causes, results from, or merely co-varies with the ageing process. CausAge applies Mendelian-randomization-informed causal inference to identify CpGs whose methylation change is more likely upstream of biological ageing, and separately derives DamAge (damage-associated accelerated ageing) and AdaptAge (adaptive response) sub-clocks. The framework suggests that clock CpGs are heterogeneous in their causal role and that mortality-associated acceleration may be driven primarily by damage-linked sites rather than adaptive ones.

DamAge and AdaptAge are causality-aware epigenetic clocks developed by Kejun Ying and colleagues in the Gladyshev lab at Harvard / Brigham and Women's Hospital (Nature Aging, 2024). Not to be confused with stress-induced "damage clocks" such as the Sinclair lab's ICE-based constructs (Yang et al., Cell 2023), which describe induced epigenomic drift rather than causality-aware methylation clocks. Standard DNA-methylation clocks (Horvath, Hannum, GrimAge) are predictive but blind to whether CpG changes are causes or downstream consequences of aging. Using epigenome-wide Mendelian randomisation against longevity and disease GWAS, the authors identified CpGs causally linked to detrimental ageing outcomes (used to build DamAge - damaging methylation changes that accelerate biological age and correlate with mortality) and CpGs causally linked to beneficial outcomes (used to build AdaptAge - protective methylation changes reflecting compensatory adaptation). Both partition the broader CausAge signal by causal direction. They are responsive to short-term interventions and improve mechanistic interpretation of biological-age studies.

The DNAm Skin & Blood clock, published by Horvath and colleagues in 2018, is an epigenetic age estimator based on 391 CpG sites selected from methylation arrays applied to skin fibroblasts and blood samples. It was developed partly to overcome the observation that the original 2013 Horvath clock systematically underestimated age in keratinocytes and fibroblasts, tissues central to studies of in-vitro reprogramming and rejuvenation. Because it was trained on a tissue type directly accessible in longevity intervention studies, it is frequently used as a readout in partial reprogramming experiments alongside the multi-tissue Horvath clock.

DNAmTL is an epigenetic estimator of telomere length derived from the methylation levels of 140 CpG sites in blood DNA, trained via elastic net regression against Southern-blot-measured leukocyte telomere length in 2,256 participants from the Women's Health Initiative and Jackson Heart Study. Unlike qPCR or flow-FISH, it does not measure telomere sequence directly; instead it captures a methylation signature co-varying with telomere attrition across the replicative history of the leukocyte compartment. DNAmTL correlates with chronological age at r ≈ −0.62 to −0.80, substantially exceeding measured leukocyte telomere length (r ≈ −0.30 to −0.40). Lu, Horvath et al. (2019, Aging) validated the estimator in 9,044 methylation arrays across seven cohorts including Framingham Heart Study and TwinsUK (mean follow-up 11.8 years): each kilobase increase in age-adjusted DNAmTL associated with HR = 0.37 for all-cause mortality (p = 2.5×10⁻²⁰), HR = 0.51 for coronary heart disease (p = 6.6×10⁻⁵), and HR = 0.27 for congestive heart failure (p = 3.6×10⁻⁶) — all stronger than measured-LTL associations. Wang et al. (2024, Clinical Epigenetics; NHANES 1999–2002) confirmed the predictive advantage over qPCR-based telomere length. Liang et al. (2024, Aging Cell; n = 2,398) showed a one-kilobase decrease in DNAmTL associates with 40% higher mortality risk (HR = 0.60) and that HIV infection links to shorter DNAmTL (β = −0.25, p = 1.5×10⁻¹²). DNAmTL requires Illumina EPIC or 450k array data; causal direction remains unresolved.

DunedinPACE (Pace of Aging Calculated from the Epigenome) is an epigenetic clock published in 2022 by Belsky and colleagues that estimates the rate of biological ageing rather than a static age. It was trained in the Dunedin 1972-1973 birth cohort on longitudinal change across 19 organ-system biomarkers and translated into a DNA-methylation score using 173 CpGs. The score is calibrated so that 1 represents the cohort-mean pace of one year of biological aging per chronological year; values above 1 indicate faster-than-average ageing. DunedinPACE shows good test-retest reliability and predicts morbidity and mortality.

Epigenetic age is a biological-age estimate derived from DNA-methylation patterns at selected CpG sites, computed by algorithms known as epigenetic clocks (e.g. Horvath, Hannum, PhenoAge, GrimAge, DunedinPACE). The difference between epigenetic and chronological age, called epigenetic age acceleration, is associated with mortality, cardiovascular disease and cancer in research cohorts. Validation depends on the specific clock; first-generation clocks track chronological age, while mortality-trained clocks better predict health outcomes. Commercial consumer tests vary in reliability.

GlycanAge is a biological-age estimate derived from the N-glycan composition of immunoglobulin G (IgG), measured in blood plasma by high-throughput capillary electrophoresis or ultra-performance liquid chromatography. IgG glycosylation shifts predictably with age—specifically, a decline in galactosylation and sialylation alongside a rise in bisecting GlcNAc accompanies the shift toward a more pro-inflammatory IgG glycome—and these patterns also respond to lifestyle interventions and chronic disease. Because glycans regulate IgG effector function and inflammaging, the measure captures an immunologically relevant dimension of ageing not directly accessible to DNA-methylation clocks; however, reference populations and clinical thresholds are still under active investigation.

GrimAge is a second-generation epigenetic clock introduced by Lu et al. (2019, with Steve Horvath as senior author). Instead of predicting chronological age, it is trained on time-to-death and combines DNA-methylation surrogates for seven plasma proteins (e.g. PAI-1, GDF-15) and DNAm-based smoking pack-years. In multiple cohorts GrimAge and the updated GrimAge2 (2022) outperform earlier clocks at predicting all-cause mortality, cardiovascular disease and cancer. It is widely used in research; clinical use as a diagnostic endpoint remains investigational.

The Hannum clock is a blood-based epigenetic age estimator published by Gregory Hannum and colleagues in 2013. It uses DNA methylation levels at 71 CpG sites, derived from whole-blood samples of 656 individuals, to predict chronological age with a cross-validated correlation of ~0.96. Unlike the Horvath clock, which is multi-tissue, the Hannum clock was trained and validated specifically in blood, making it less generalisable to other tissues. It remains widely cited but has been largely superseded for mortality prediction by second-generation clocks trained on health outcomes.

The Horvath clock is a multi-tissue epigenetic age estimator published by Steve Horvath in 2013. It uses DNA methylation levels at 353 CpG sites to predict chronological age across more than 50 tissues and cell types with a median error of about 3.6 years. It is the most-cited epigenetic clock and well validated as a predictor of chronological age, but its association with mortality and disease is weaker than that of later, mortality-trained clocks such as GrimAge.

iAge is an inflammatory-age metric introduced by Sayed and colleagues (2021, Stanford) that uses a deep-learning model trained on a panel of 50 circulating cytokines and chemokines from 1,001 healthy individuals across the decade-long Stanford 1000 Immunomes Project. The model compresses the immunome profile into a single inflammatory-age score that predicts cardiovascular risk, multimorbidity, and all-cause mortality independently of chronological age. CXCL9—a chemokine associated with T-cell recruitment and endothelial dysfunction—was identified as the single most informative driver. The clock highlights the immune system as a distinct, targetable dimension of biological ageing.

The Klemera-Doubal method (KDM) is a statistical algorithm that estimates biological age from clinical biomarkers by minimizing the sum of squared distances between m regression lines and m biomarker values in multidimensional space. Each biomarker is first regressed on chronological age in a reference population, yielding slope hᵢ, intercept gᵢ, and root mean squared error sᵢ; these per-biomarker weights are pooled with a chronological-age anchor (variance sD²) into a single KDM-BA score. The key architectural departure from multiple linear regression (MLR) is that biomarkers regress onto age rather than age onto biomarkers, which reduces error propagation and collinearity. Levine (2013, J Gerontol A) tested five biological age algorithms in 9,389 NHANES III participants followed 18 years (1,843 deaths): KDM-BA reached an AUC of 0.851 vs 0.827 for chronological age alone, hazard ratio 1.09 per year (95% CI 1.08–1.09); chronological age became non-significant when combined with KDM-BA, a property MLR scores did not match. KDM-BA then served as the training target in Levine et al. (2018): a Gompertz proportional-hazards model converted NHANES III 10-year mortality risk into phenotypic age (PhenoAge) from nine blood biomarkers, the clinical precursor step to DNAm PhenoAge. Both measures are standard reference algorithms in the BioAge R toolkit (Kwon and Belsky, 2021, GeroScience). Homeostatic dysregulation scores and machine-learning clocks have since outperformed KDM in head-to-head mortality comparisons, positioning it as a validated, interpretable benchmark as of 2026.

Leukocyte telomere length (LTL) is the average length of repetitive TTAGGG sequences capping chromosome ends in white blood cells, measured in kilobases (kb), used as a proxy for cumulative replicative stress. Two assay platforms exist: quantitative PCR (qPCR), cost-effective for large cohorts, and flow-FISH, which achieves lower inter-assay variability (CV ~10% vs ~16% for qPCR). LTL declines roughly 20-40 bp per year in adults; critically short telomeres activate p53-mediated senescence. In 422,797 UK Biobank participants, Bountziouka et al. (2022) identified smoking and brisk walking pace as the strongest modifiable correlates, though all tested traits combined explained less than 0.2% of variance. Shorter LTL correlates with higher cardiovascular and all-cause mortality risk at population level, but individual predictive value is limited by wide distributional overlap across age groups. Mendelian randomisation yields a heterogeneous picture: genetically shorter LTL raises risk for cardiovascular disease and multiple sclerosis, while longer LTL paradoxically increases risk for soft-tissue sarcoma and atrial fibrillation. Direct-to-consumer tests lack standardised reference ranges and explain only a modest fraction of biological-age variance relative to epigenetic clocks.

OMICmAge is a DNA-methylation-based biological age clock that incorporates information from the proteome, metabolome, and routine clinical laboratory data into a single blood-derived estimate — without requiring direct measurement of those additional layers at the time of testing. The clock was developed by Chen, Dwaraka, Carreras-Gallo, Higgins-Chen, Lasky-Su and colleagues, published in Nature Aging (2026), building on a bioRxiv preprint from 2023. Its construction begins with EMRAge, a mortality-linked composite derived from 19 clinical laboratory values across approximately 31,000 Mass General Brigham Biobank records; elastic-net regression then models EMRAge from a candidate pool of 396 epigenetic biomarker proxies (EBPs) — methylation-based surrogates for 266 metabolites, 109 proteins, and 21 clinical variables — and retains 40 EBPs (16 protein, 14 metabolite, 10 clinical) alongside 990 CpG (cytosine-phosphate-guanine) sites in the final model, distilling multi-omic information into an output readable from DNA methylation alone. Validation across independent cohorts from TruDiagnostic (n = 14,213) and Generation Scotland (n = 18,672) showed OMICmAge acceleration associated with type-2 diabetes, stroke, cardiovascular disease, COPD, depression, and cancer, and yielded 5- and 10-year mortality AUC values of approximately 0.89 and 0.87 — several percentage points above PCGrimAge or chronological age alone in head-to-head comparisons. OMICmAge is currently a research tool; it has not been cleared by any regulatory authority as a diagnostic, and its relationship to the aging process remains associational.

PCGrimAge applies the same principal-component denoising framework introduced by Higgins-Chen et al. (2022) to the GrimAge methylation features, producing a more technically stable version of one of the strongest mortality-predictive epigenetic clocks. By regressing out array-platform and batch variance before scoring, PCGrimAge shows markedly better within-individual reproducibility than the original GrimAge algorithm (Lu et al. 2019), an important property when tracking biological-age change in response to interventions such as caloric restriction, exercise, or pharmacological treatments. Like PCPhenoAge, it represents a methodological advance rather than a new biological model.

PCPhenoAge is a technically refined variant of DNAm PhenoAge introduced by Higgins-Chen and colleagues (2022) that applies principal-component (PC) regression to the underlying CpG data before computing the age score. The PC transformation removes technical noise and batch effects that inflate variance in standard methylation arrays, yielding a clock with substantially improved test-retest reliability and reduced sensitivity to sample quality or preprocessing pipeline. In practice, PCPhenoAge retains the mortality-predictive validity of its predecessor while being more suitable for longitudinal studies and intervention trials where within-individual change is the primary outcome.

PhenoAge is a composite biological-age measure developed by Levine and colleagues in 2018. The original blood-based version combines nine clinical biomarkers including albumin, creatinine, glucose, C-reactive protein and white blood cell count with chronological age, calibrated against mortality. A DNA-methylation version, DNAm PhenoAge, transfers the score to epigenetic data. PhenoAge predicts all-cause mortality and multimorbidity better than chronological age and has been validated in several large cohorts, although clinical use is still emerging.

ProAge and related proteomic-age clocks estimate biological age from the concentrations of hundreds to thousands of plasma or serum proteins measured by aptamer-based (SomaScan) or proximity-extension assay (Olink) platforms. Landmark studies by Lehallier and colleagues (2019, Nature Medicine) demonstrated that the plasma proteome changes non-linearly with age in three distinct waves, and subsequent work trained predictive models on up to ~3,000 proteins. Proteomic clocks capture post-transcriptional and secreted signals not reflected in DNA methylation, and recent analyses suggest protein-based age acceleration associates with age-related disease risk, though platform-specific protein selection means scores are not directly interchangeable across studies.

RetinaAge is a biological-age clock derived from fundus photographs of the retina, using deep-learning models trained to predict chronological age from the vascular and neural features of the optic disc, fovea, and retinal vasculature. Published by Zhu and colleagues (2022, British Journal of Ophthalmology), the model was trained on fundus images from healthy UK Biobank participants and applied to the broader UK Biobank cohort (~80,000 images, ~46,000 participants); the gap between predicted retinal age and chronological age independently predicted all-cause mortality; a separate analysis from the same group (Zhu et al., 2022, Stroke) linked the gap to arterial stiffness and incident cardiovascular disease. Because the retina shares embryological origin with the central nervous system and is the only site where microvasculature can be imaged non-invasively, it offers a clinically practical window onto systemic and neural vascular aging.

SystemsAge is a multi-organ biological-age framework introduced by Tian and colleagues (2023, Nature Medicine) that uses longitudinal brain MRI and physiological/biomarker phenotypes from large biobanks (UK Biobank in the original paper) to generate organ-specific age estimates across 3 brain systems (cerebral cortex, thalamus, cerebellum) and 7 body systems (cardiovascular, respiratory, renal, gastrointestinal, liver, musculoskeletal, immune) — 10 systems total. Rather than a single composite score, it produces a profile of organ biological ages, allowing detection of discordant ageing across systems. In the UK Biobank, organ age gaps predicted organ-specific disease and all-cause mortality; individuals whose organ ages were globally younger than their chronological age showed lower mortality risk. A separate methylation-based 'SystemsAge' clock by Sehgal et al. 2025 Nature Aging (Levine lab) covers 11 physiological systems via a single blood-based assay; the two clocks share a name but use different inputs.