What David Sinclair's Biological Age Test Measures

Woolf Software
epigenetics biological-age methylation biomarkers measurement
A feathered reptile-like creature on a black plinth, its spine and flanks studded with rows of tiny beads glowing at varying brightness.

When people search for “David Sinclair biological age test,” the product they have in mind is TallyAge. It is sold by Tally Health, the company Sinclair co-founded. It works from a cheek swab, measuring DNA methylation at a set of CpG sites in buccal epithelial cells, and it returns a single number in years alongside a subscription and a supplement line. The method underneath is an epigenetic clock, meaning a penalized regression model trained to predict either chronological age or a mortality-linked phenotype from methylation beta values. This is a genuine measurement class with a large scientific literature behind it. The difficulties arise in how it is packaged for consumers. They come down to three things: precision, the choice of tissue, and the fact that you never see the raw data.

What an epigenetic clock is, mechanically

Before judging any commercial test, it helps to understand what the underlying measurement is and what the various clocks were built to do. Methylation at a CpG site is reported as a fraction between 0 and 1, representing the proportion of cells in your sample that carry a methyl group at that position. Hundreds of thousands of those fractions are measured at once on an Illumina array, and a clock is simply a linear model built over a subset of them.

The published clocks differ in how many sites they use and what they were trained against. Horvath’s 2013 multi-tissue clock uses 353 CpGs and predicts chronological age with a median absolute error of about 3.6 years across many tissue types.1 Hannum’s blood clock uses 71. PhenoAge uses 513 and is trained against a mortality-weighted composite of clinical labs rather than against age itself. GrimAge takes a two-stage approach: it first builds DNA methylation surrogates for seven plasma proteins and for smoking pack-years, then runs a mortality regression over those surrogates.

The generational split matters more than the site counts. First-generation clocks were trained on chronological age, so their ceiling is predicting something you already know from your birth certificate. Second-generation clocks were trained on outcomes instead, and they perform better on the outcomes you care about. In a head-to-head comparison across several clocks in an aging cohort, GrimAge was the strongest predictor of all-cause mortality and age-related clinical phenotypes.2 The quantity that carries predictive value for disease and mortality in older adults is epigenetic age acceleration, the residual left after regressing methylation age on chronological age.3 A review of epigenetic aging for population health sets out how these measures are constructed and where they are and are not fit for individual-level use.4

Sinclair’s own scientific contributions sit upstream of the consumer test rather than inside it. His lab’s framing is that methylation change is not only a readout of aging but part of its mechanism, a loss of epigenetic information that can be partially restored in mice.5 The information-theoretic version of that argument is laid out in his 2023 paper with Lu and Tian.6 His group also worked on the economics of the assay itself. TIME-seq uses transposase-based library preparation with early pooling to cut the cost of methylation-based age prediction by roughly two orders of magnitude compared with arrays.78 And his group has shown that multi-omic clocks combining methylation with other molecular layers predict phenotypic age in mice better than methylation alone.9

Why the number moves when nothing about you has

A single consumer result carries more variance than the clean presentation suggests, and three sources dominate. Understanding them is what keeps you from reacting to a change that never happened.

The first is technical noise. Array clocks are sums over individual CpGs, each of which is measured with error. Run the same DNA twice on two chips and first-generation clock outputs can differ by a year or more. Consumer reports rarely show a confidence interval, so a difference of two years between a January and a July sample sits inside the noise band for most single-run assays. If a test ships you one number with no error bar, the trailing digit is decoration.

The second is cell composition. A cheek swab collects a mixture of buccal epithelial cells and leukocytes. The ratio varies with how hard you scraped, how long it has been since you ate, and whether you have any oral inflammation. Methylation differs sharply between those cell types, so a shift in the mixture moves the clock even though nothing in your biology has changed. Blood has the same problem at a smaller amplitude, where the neutrophil-to-lymphocyte ratio drives apparent age acceleration. That is why careful blood analyses regress out estimated cell proportions, using Houseman deconvolution or EpiDISH in R, before interpreting anything.

The third is training-set mismatch. Most well-validated clocks were trained on blood or on multi-tissue panels. A buccal-trained proprietary model may be perfectly self-consistent and still not be comparable to anything in the published literature. You cannot put a TallyAge number next to a published GrimAge acceleration and draw a conclusion.

What we would do instead

Our recommendation is straightforward: buy the raw data rather than the number. Owning the underlying measurements lets you check quality, compute published clocks, and compare your own results to each other over time.

In practice, that means ordering a methylation array run on whole blood through a laboratory that returns IDAT files, the per-sample raw intensity files the Illumina scanner produces. Ask for Illumina EPIC v2 (about 935k probes) or EPIC v1 (about 850k). Insist on the red and green IDATs for each sample plus the sample sheet listing sentrix ID and position, because chip position is a real batch covariate and you will want it later. From there, the processing happens locally, and the steps are these:

  • Load the data with sesame (our preference) or minfi in R/Bioconductor. A single call to openSesame(idat_dir, prep="QCDPB") returns detection-masked, background-corrected, dye-bias-corrected beta values.
  • Check quality before doing anything else. Look for a mean detection p-value per sample below 0.01 and confirm the bisulfite conversion controls. Then compare predicted sex against known sex, and check predicted SNP genotype concordance across your own timepoints. The array’s 59 rs probes act as a fingerprint and will catch a sample swap at the lab.
  • Compute clocks with methylclock, or use Horvath’s calculator input format, which is a CSV of probe IDs by samples plus an annotation file with age and tissue. Expect some probes to be missing on EPIC v2 relative to the 450k-era clock definitions. Impute those with the training-set mean rather than dropping them, and record how many probes you imputed. If more than a few percent of a clock’s CpGs are missing, that clock’s output is not trustworthy for you.
  • Estimate cell fractions with EpiDISH (RPC mode, blood reference) and keep them alongside every result.

Then comes the step most people skip, which is running technical replicates. Use two aliquots of the same blood draw, on the same chip, in different positions. That gives you a measured noise floor for your own pipeline, and any difference smaller than that floor is not a change. Sampling every six to twelve months under standardized conditions gives you a slope. Standardized means the same time of day, the same fasting state, and the same lab. The slope is the interpretable quantity. A single absolute number is nearly uninterpretable at the individual level.4

The blood-panel alternative you can compute from labs you may already have

If a methylation array feels like a large first step, there is a cheaper measure you can calculate from routine bloodwork. PhenoAge’s clinical version needs nine lab values plus chronological age: albumin, creatinine, glucose, log C-reactive protein, lymphocyte percent, mean cell volume, red cell distribution width, alkaline phosphatase, and white blood cell count. A CBC with differential and a comprehensive metabolic panel with hs-CRP together cover all of them.

The coefficients are published in the original PhenoAge paper, and the arithmetic amounts to about fifteen lines of Python. The measurement costs a tiny fraction of a methylation array, repeats cheaply, and its inputs are quantities a clinician can act on directly if one of them falls out of range. For those reasons we treat PhenoAge as the first biological age measure to compute and methylation as the second.

Neither number is a diagnosis. An elevated age acceleration is a statistical statement about a population rather than a finding about your body. The specific labs that feed PhenoAge, including CRP, RDW, creatinine, and glucose, are ones where an abnormal value should go to a physician rather than into a spreadsheet.

Questions people also ask

What is the most accurate test for biological age? Accuracy depends on what you are trying to predict. For predicting chronological age, first-generation clocks land within a few years.1 For predicting mortality and age-related clinical phenotypes, GrimAge outperformed the other clocks tested in a direct comparison.2 For tracking yourself over time, precision matters more than accuracy, which means using the same assay on the same tissue, replicated, over a period of years.

How do I know my real biological age? There is no single ground truth to check a result against. What exists is a set of models whose age-acceleration residuals correlate with disease and mortality risk in cohorts.3 The defensible reading of a result is that your methylation pattern resembles that of people somewhat older or younger than you, and the defensible use of it is watching the trend across your own repeated measurements.

What is David Sinclair’s anti-aging routine? He has described his personal regimen in public interviews and in his book, and it has included supplements and a prescription drug. We do not evaluate or recommend any of it. Nothing in it is established to change human lifespan, and anything prescription belongs in a conversation with your physician.

What is the most powerful anti-aging supplement? No supplement has been shown to reduce all-cause mortality or slow a validated aging biomarker in a well-powered human trial. Sinclair’s lab work on epigenetic reprogramming is mouse work and mechanism work.59 The sensible posture is to treat the measurement side as tractable today and the intervention side as unsettled.

Woolf Software builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

  1. Steve Horvath. DNA methylation age of human tissues and cell types. Genome Biology, 2013. https://doi.org/10.1186/gb-2013-14-10-r115 ↩ ↩2

  2. Cathal McCrory, Giovanni Fiorito, Belinda Hernandez, et al. GrimAge Outperforms Other Epigenetic Clocks in the Prediction of Age-Related Clinical Phenotypes and All-Cause Mortality. The Journals of Gerontology: Series A, 2020. https://doi.org/10.1093/gerona/glaa286 ↩ ↩2

  3. Qihua Tan. Epigenetic age acceleration as an effective predictor of diseases and mortality in the elderly. EBioMedicine, 2021. https://doi.org/10.1016/j.ebiom.2020.103174 ↩ ↩2

  4. Cynthia D.J. Kusters, Steve Horvath. Quantification of Epigenetic Aging in Public Health. Annual Review of Public Health, 2025. https://doi.org/10.1146/annurev-publhealth-060222-015657 ↩ ↩2

  5. Alice E. Kane, David A. Sinclair. Epigenetic changes during aging and their reprogramming potential. Critical Reviews in Biochemistry and Molecular Biology, 2019. https://doi.org/10.1080/10409238.2019.1570075 ↩ ↩2

  6. Yuancheng Ryan Lu, Xiao Tian, David A. Sinclair. The Information Theory of Aging. Nature Aging, 2023. https://doi.org/10.1038/s43587-023-00527-6 ↩

  7. Patrick T. Griffin, Alice E. Kane, Alexandre Trapp, et al. TIME-seq reduces time and cost of DNA methylation measurement for epigenetic clock construction. Nature Aging, 2024. https://doi.org/10.1038/s43587-023-00555-2 ↩

  8. Patrick T Griffin, Alice E Kane, Alexandre Trapp, et al. TIME-Seq Enables Scalable and Inexpensive Epigenetic Age Predictions. 2021. https://doi.org/10.1101/2021.10.25.465725 ↩

  9. Daniel L Vera, Patrick T Griffin, David Leigh, et al. Multiomic clocks to predict phenotypic age in mice. The Journals of Gerontology, Series A: Biological Sciences and Medical Sciences, 2025. https://doi.org/10.1093/gerona/glaf188 ↩ ↩2