Methods and Data for Aging Clocks

Clinical and functional composites

Combine routine chemistry, hematology, blood pressure, lung or kidney function, cognition, frailty, or other physiological measures using regression, distance, or survival models.

• Representative methods: Klemera-Doubal biological age, clinical Phenotypic Age, homeostatic dysregulation, frailty indices.
• Strength: direct connection to function and large epidemiological cohorts.
• Limit: acute illness, medication, fasting state, and laboratory platform can shift scores.

DNA methylation clocks

Use methylation at selected CpGs or genome-wide patterns with penalized regression, principal components, survival training, or deep learning.

• Representative models: Hannum, Horvath pan-tissue, Skin & Blood, DNAm PhenoAge, GrimAge, DunedinPACE, AltumAge.
• Strength: mature assays, published coefficients, and broad benchmark literature.
• Limit: probe coverage, normalization, tissue, blood-cell composition, batch, ancestry, and platform version affect results.

Transcriptomic clocks

Learn age-associated gene-expression patterns from bulk RNA-seq, microarrays, pseudobulk profiles, or cell-type-specific expression.

• Representative models: Peters whole-blood transcriptional age and RNAAgeCalc-style tissue models.
• Strength: connects age estimates to active pathways and cell states.
• Limit: RNA quality, ischemic time, library chemistry, circadian state, inflammation, and composition are major sources of drift.

Proteomic and metabolomic clocks

Use circulating proteins, lipids, metabolites, glycans, or composite molecular panels to estimate age or age-related risk.

• Representative approaches: SomaScan/Olink proteomic age, plasma proteomic aging signatures, NMR MetaboAge-type models.
• Strength: sensitive to systemic physiology and potentially modifiable exposures.
• Limit: platform-specific feature definitions and calibration make coefficient transfer difficult without bridging data.

Imaging and physiological clocks

Apply machine learning to MRI, retinal images, ECG, wearable signals, facial images, or organ-function measurements.

• Representative approaches: T1-MRI brain age, retinal age gap, ECG age, organ-specific age models.
• Strength: spatial or functional readouts linked to organ integrity.
• Limit: scanner or device domain shift, acquisition protocol, disease prevalence, and site leakage can dominate apparent accuracy.

Single-cell and multi-omics clocks

Estimate age from single-cell methylomes, cell-type-resolved expression, chromatin, or integrated molecular layers.

• Representative approaches: scAge for single-cell DNA methylation, donor-level cell-type clocks, multi-view latent models.
• Strength: separates cell-intrinsic changes from tissue composition and identifies asynchronous cell aging.
• Limit: sparse measurements, uneven cells per donor, batch, cell-state imbalance, and small donor counts create severe overfitting risk.

NCBI Gene Expression Omnibus

Public functional-genomics studies spanning methylation arrays, expression arrays, bulk RNA-seq, and many single-cell datasets.

• Useful examples: GSE40279 and GSE87571 for age-stratified whole-blood methylation.
• Audit first: sample reuse, processed-versus-raw values, age coding, platform, disease exclusions, and batch structure.

GTEx Portal

Multi-tissue human transcriptomes and genotypes with donor attributes suitable for tissue-specific age-pattern discovery.

• Best for: cross-tissue transcriptomic comparisons and external testing of tissue models.
• Audit first: age bins rather than exact age in open data, post-mortem interval, ischemic time, tissue quality, and controlled-access fields.

NHANES

Population-based clinical, laboratory, examination, exposure, and linked outcome data widely used for phenotypic-age analyses.

• Best for: clinical composites, population calibration, mortality association, and survey-weighted analyses.
• Audit first: survey cycle, assay drift, fasting subsample, missingness, weights, and linkage eligibility.

UK Biobank

Large-scale genetics, clinical phenotypes, imaging, biomarkers, accelerometry, proteomics, and longitudinal outcomes under approved access.

• Best for: multimodal clocks, organ-age models, prospective outcome validation, and subgroup analysis.
• Audit first: healthy-volunteer selection, repeat-assessment subset, ancestry representation, leakage, and data-use scope.

HRS and ADNI

Longitudinal cohorts with aging-related phenotypes; HRS includes population-based social, clinical, genomic, and methylation resources, while ADNI is enriched for cognitive aging and neurodegeneration.

• Best for: prospective outcomes, repeated measures, cognition, and external validation of established clocks.
• Audit first: access tier, disease enrichment, sampling weights, visit timing, repeated samples, and outcome definitions.

Human Cell Atlas and CELLxGENE

Curated single-cell collections for donor-aware, cell-type-resolved age analyses across tissues and studies.

• Best for: cell-type-specific hypotheses, pseudobulk models, and cross-study validation.
• Audit first: donor count, age distribution, tissue handling, cell ontology, disease, study effects, and whether exact age is available.

All of Us Research Program

A diverse longitudinal resource combining electronic health records, surveys, physical measurements, wearables, and genomic data in a secure cloud workbench.

• Best for: population-scale clinical and multimodal clocks, diversity-aware calibration, and prospective health outcomes.
• Access: aggregate Public Tier data are open; participant-level Registered and Controlled Tier data require a data passport and authorized Researcher Workbench use.

NCI Genomic Data Commons

Harmonized cancer genomic, transcriptomic, methylation, clinical, and biospecimen data from TCGA and other NCI programs.

• Best for: stress-testing tissue and disease specificity, tumor-versus-adjacent comparisons, and examining how disease distorts age predictors.
• Access: high-level derived, clinical, and biospecimen data are often open; raw sequence, germline, genotype, and protected files require eRA Commons and project-specific dbGaP authorization.

Dunedin Study

A deeply phenotyped 1972-1973 New Zealand birth cohort with repeated assessments across the life course; its longitudinal organ-system changes underpin the Dunedin pace-of-aging measures.

• Best for: longitudinal pace targets, within-person change, life-course exposure analyses, and multimodal replication.
• Access: participant-level data are not public. Investigators use a managed process involving a concept paper, study collaboration, review, and secure analysis under the Study policy.

Lock the estimand and cohort

• 1. Predefine the target, output unit, intended use, inclusion criteria, and analysis time point.
• 2. Split by donor and, for transport tests, by study or site; keep repeats, tissues, and all cells from one donor in the same partition.

Reproduce the input pipeline

• 3. Match tissue, assay, feature identifiers, normalization, transformation, missing-feature handling, and model version.
• 4. Quantify sample QC, probe or feature coverage, technical replicate reliability, batch effects, and limits of detection.

Report more than correlation

• 5. For age prediction report MAE/RMSE, correlation, calibration slope/intercept, residual plots, and uncertainty by age band.
• 6. For risk or pace models report suitable survival/longitudinal metrics and compare against age, sex, and established baselines.

Test transportability and fairness

• 7. Validate in an independent cohort and report performance by sex, ancestry, age range, tissue, disease state, platform, and site where powered.
• 8. Model key composition and exposure variables such as blood-cell fractions, smoking, medication, BMI, and inflammation; state whether adjustment removes part of the intended signal.

Establish biological and clinical validity

• 9. Test prospective outcomes, within-person change, convergent and discriminant validity, and incremental value beyond chronological age.
• 10. For interventions require randomization where feasible, blinded processing, target engagement, prespecified time points, durability, functional outcomes, and adverse-event assessment.

Version and release the analysis trail

• Record code commit, coefficient source, model version, package/container, reference data, licenses, random seeds, and all preprocessing parameters.
• Publish a model card with training population, intended use, excluded uses, known biases, missing-data policy, and validation results.