Biolearn Benchmarks 39 Aging Clocks and Shows Age Accuracy Does Not Predict Mortality
An open-source tool tested 39 aging biomarkers across 20,000+ people. Clocks best at predicting age were not the best at predicting death or disease.
Summary
Researchers built Biolearn, an open-source Python library that harmonizes 39 aging biomarkers (epigenetic, transcriptomic, and clinical) and the datasets used to test them. Benchmarking across many cohorts showed that how well a clock predicts chronological age is unrelated to how well it predicts mortality after adjusting for age and sex. The Horvath skin-and-blood clock was most accurate for age (R2 = 0.88), while GrimAge2 was the strongest mortality predictor in the largest cohort (HR = 2.57 per SD) and for healthspan (HR = 2.00). Performance varied by outcome: DunedinPACE led for stroke and dementia, and PhenoAge for metabolic disease. The framework also provides quality control, imputation, and blood cell-type deconvolution, supporting reproducible, community-driven biomarker validation.
Detailed Summary
Aging clocks are now widely used to estimate biological age, but comparing them is hard. Each uses different input formats, tissues, and training data, and validation datasets are fragmented. This makes it difficult to know which biomarker is best for which purpose, a key hurdle before clinical use.
To address this, the team built Biolearn, an open-source Python library. It standardizes 39 published biomarkers (DNA methylation clocks such as Horvath, Hannum, PhenoAge, GrimAge, GrimAge2, DunedinPACE, and causality-enriched clocks, plus transcriptomic tAge predictors and the blood-chemistry Phenotypic Age). Implementations were verified with the original developers. Biolearn offers about 2,000 pre-processed datasets, including 40 manually curated references, and underpins ClockBase's processing of more than 200,000 GEO samples. It includes quality control, sex/smoking/BMI prediction, imputation, and deconvolution of six blood cell types. In a test where 20% of methylation values were removed, imputation restored Horvath age predictions (R from 0.853 to 0.953; MAE from 8.96 to 2.40 years).
For chronological age, eight public cohorts showed the Horvath skin-and-blood clock was most accurate (mean R2 = 0.88), followed by Hannum (0.81), Horvath multi-tissue (0.78), and Ying CausAge (0.77). For mortality, the team assessed 14 clocks in three cohorts: Normative Aging Study (N = 1,488), a Massachusetts General Brigham cohort (N = 500), and Generation Scotland (N = 18,859). GrimAge2 was top or near-top in each (HR 1.35, 2.08, and 2.57 per SD, respectively), though DunedinPoAm38 and DunedinPACE were comparable in NAS. Critically, accuracy for chronological age did not correlate with mortality prediction (R = 0.12, P = 0.67).
Outcome-specific results differed. In Generation Scotland, GrimAge2 was strongest for healthspan (HR 2.00), cardiovascular disease (1.70), and respiratory disease (4.38), PhenoAge for metabolic conditions (2.28), and Horvath skin-and-blood for cancer (1.73). In NAS, DunedinPACE led for stroke, dementia, and healthspan. Healthspan and lifespan associations were highly correlated across clocks. Clocks clustered into two groups, with telomere-length and AdaptAge clocks distinct. tAge predictors correlated with age (R 0.68–0.90) in a small RNA-seq set, and blood cell proportions linked to outcomes: higher NK-cell fraction was protective, while neutrophils and CD8 T cells were risk-associated.
Implications: biomarker choice should match the question, and age deviation should be interpreted cautiously. Caveats include training-data overlap (e.g., Framingham, NHANES), mostly linear models, dataset-to-dataset variability, and retrospective NAS analyses.
Key Findings
- Chronological-age accuracy of clocks did not correlate with mortality prediction (R = 0.12, P = 0.67), suggesting they capture different biology.
- Horvath skin-and-blood clock was most accurate for age (mean R2 = 0.88) across eight cohorts, ahead of Hannum (0.81).
- GrimAge2 had the strongest mortality association in Generation Scotland (HR 2.57 per SD, N = 18,859) and healthspan (HR 2.00).
- Best biomarker varied by outcome: DunedinPACE for stroke and dementia, PhenoAge for metabolic disease, Horvath v2 for cancer.
- Imputation restored Horvath predictions after 20% simulated missingness (MAE 8.96 to 2.40 years); deconvolution matched FACS-measured blood cell proportions.
Methodology
Biolearn, an open-source Python library, harmonized 39 biomarkers and public datasets (GEO, NHANES, Framingham-related). Clock accuracy was benchmarked in eight cohorts. Fourteen clocks were tested for mortality and disease using age- and sex-adjusted Cox models in NAS, MGB, and Generation Scotland, with hazard ratios per SD.
Study Limitations
Several clocks (DunedinPACE, GrimAge, PhenoAge) were trained on datasets that overlap with some validation data, which can inflate performance. Biolearn mainly implements linear models, and performance varied across cohorts because of batch effects, demographics, and disease states. The NAS analyses were retrospective, event rates in some disease categories were low, and the transcriptomic evaluation used only 62 samples.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: