<?xml version="1.0" encoding="UTF-8"?>
<feed xmlns="http://www.w3.org/2005/Atom" xml:lang="en">
    <title>Bio Clock Test: biological age tests, read carefully</title>
    <subtitle>Biological age tests explained from the research: epigenetic clocks like Horvath, GrimAge and DunedinPACE, blood-marker clocks, consumer kits, how reliable a result is and what actually changes it.</subtitle>
    <link rel="self" type="application/atom+xml" href="https://bioclocktest.com/atom.xml"/>
    <link rel="alternate" type="text/html" href="https://bioclocktest.com"/>
    <generator uri="https://www.getzola.org/">Zola</generator>
    <updated>2026-11-01T00:00:00+00:00</updated>
    <id>https://bioclocktest.com/atom.xml</id>
    <entry xml:lang="en">
        <title>Biological Age Calculators From Your Blood Work</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/biological-age-calculator/"/>
        <id>https://bioclocktest.com/biological-age-calculator/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/biological-age-calculator/">&lt;p&gt;There is a formula that takes nine numbers from an ordinary blood test, the kind that comes back as a PDF full of reference ranges, and turns them into an age. It is called PhenoAge, and its nine inputs include creatinine, blood glucose, C-reactive protein and lymphocyte percentage, which is the share of your white blood cells that happen to be lymphocytes. You put in chemistry. Out comes a number with “years” after it.&lt;&#x2F;p&gt;
&lt;p&gt;That is what a biological age calculator built on routine labs is: a weighted formula, fitted to data from a large population, that takes values a doctor may already have on file and reports one age instead of a page of results. Different calculators use different lists. One company’s series names fasting glucose, CRP, albumin, alkaline phosphatase, creatinine, red cell distribution width, mean corpuscular volume, lymphocytes and white blood cells. InsideTracker’s version, called InnerAge 2.0, includes glucose and GGT, a liver enzyme. What the facts do not show is any agreed standard that would let you check one calculator’s answer against another’s (this turns out to matter quite a bit).&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-a-biological-age-calculator-does-with-your-numbers&quot;&gt;What a biological age calculator does with your numbers&lt;&#x2F;h2&gt;
&lt;p&gt;The basic move is weighting. You take a lot of people, you know their real ages, you measure their blood, and you look for the blood numbers that move with age. The ones that move a lot get a big weight. The ones that move a little get a small one. Then you add it all up.&lt;&#x2F;p&gt;
&lt;p&gt;InsideTracker is unusually plain about this. Its calculation gives glucose “a higher weight than GGT because it is more strongly correlated with age in InsideTracker’s dataset.” Notice what that sentence does not say. It does not say glucose is the cause of aging, or that GGT is unimportant to your liver. It says that in one company’s pile of customers, glucose tracked the calendar more closely.&lt;&#x2F;p&gt;
&lt;p&gt;It is a GPA, basically. Glucose is a four-credit lab course and GGT is a one-credit seminar that meets on Fridays, and your final number depends on how the registrar assigned the credits. Or, if you prefer sports, it is the NFL passer rating for your bloodstream: a formula almost nobody can compute in their head, built from real stats, that produces a single number people then argue about at dinner.&lt;&#x2F;p&gt;
&lt;p&gt;The company also describes InnerAge 2.0 as using “clinically validated blood biomarkers.” Translated into one flat sentence: the numbers come from standard tests that labs already know how to run. That is a real virtue (it is cheap and familiar) but it is a claim about the ingredients, not about the recipe.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;starting-with-one-number-and-adding-more&quot;&gt;Starting with one number and adding more&lt;&#x2F;h2&gt;
&lt;p&gt;Here is a hypothetical, and I am making it up. Start small and reasonable. You notice that, across a big population, older people tend to have higher fasting glucose. So you build a calculator with one input: higher glucose, older age. This is dumb, but not crazy. It is the sort of thing a first-year statistics student would do on a napkin.&lt;&#x2F;p&gt;
&lt;p&gt;Then you add a second input, GGT, with a smaller weight because it tracks age less closely. Now it is a little less dumb. Then you go to nine inputs: albumin, creatinine, CRP, white cell counts, the red blood cell stuff. Each one picks up some piece of the age signal the others missed.&lt;&#x2F;p&gt;
&lt;p&gt;Then you think, why only blood? Add urine. Add a physical exam. Add behavior. Add income and education, because those correlate with health too. Then you stop choosing the weights by hand and let a deep learning model pick them. Then, and this is where it gets weird, you train a second calculator, using a totally different kind of measurement, whose whole job is to predict the output of the first calculator. Then a third, smaller calculator whose job is to match the first one using fewer inputs. Then you sell the result, split into nine organ systems, with a nice app.&lt;&#x2F;p&gt;
&lt;p&gt;The absurd endpoint is a number that is an estimate of an estimate of a weighted average of lab values, reported to you as how old your kidneys are.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-far-along-that-road-the-real-world-already-is&quot;&gt;How far along that road the real world already is&lt;&#x2F;h2&gt;
&lt;p&gt;Pretty far. Every step above exists.&lt;&#x2F;p&gt;
&lt;p&gt;A study using NHANES, the big federal health survey, built biological ages from 30 features chosen by a method called LASSO (a statistical technique that, given a huge spreadsheet, keeps the columns that help and zeroes out the rest). The 30 features covered blood and urine biochemistry, physical examination data, behavioral traits and socioeconomic factors. So the income step is not hypothetical. It has been done.&lt;&#x2F;p&gt;
&lt;p&gt;The deep learning step has been done too, at least in mice. A hematology clock built from routine blood markers in the Study of Longitudinal Aging in Mice found that platelets were the strongest predictor of biological age, and that a bigger aging gap went along with higher mortality risk and frailty. The mice did not consent to the app.&lt;&#x2F;p&gt;
&lt;p&gt;The calculator that predicts a calculator is DNAm PhenoAge. Morgan Levine identified DNA methylation sites correlated with nine clinical chemistry biomarkers of biological age, and with chronological age, using NHANES data. So the blood-chemistry formula became the target, and the methylation formula learned to hit it. Elysium Health’s product, Index, uses DNAm PhenoAge as the basis of its biological age, and it reports separate ages for nine body systems: brain, liver, kidney, immune system, heart, blood, metabolic, inflammation and hormone. That is the nine-organ step.&lt;&#x2F;p&gt;
&lt;p&gt;And the smaller calculator that matches the bigger one is GeroAge. It needs six blood markers (albumin, log CRP, lymphocyte percentage, creatinine kinase, blood glucose and white blood cell count) plus chronological age, and it shows R=0.99 agreement with PhenoAge, with similar predictive power for overall survival in NHANES data. An R of 0.99 is, in the technical sense, close. Six inputs doing the work of nine is a perfectly sensible engineering result. It also means one formula’s main credential is that it agrees with another formula.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-the-inputs-actually-mean&quot;&gt;What the inputs actually mean&lt;&#x2F;h2&gt;
&lt;p&gt;A common way to group the markers is by what system they report on: metabolic health (glucose, LDL cholesterol, triglycerides), sex hormones (testosterone, SHBG, DHEAS), liver health (GGT) and immune markers (white blood cells). Other calculators add albumin, alkaline phosphatase, creatinine and the two red blood cell measures, red cell distribution width and mean corpuscular volume, whose names describe what they count.&lt;&#x2F;p&gt;
&lt;p&gt;But inside a calculator, an input means something narrower than it does on your lab report. It means: in this training population, this number moved with age (or with death, or with a composite score) by this much. I feel like the actual conversation at a company goes something like, “Hey, glucose is tracking age really well in our customer data, can we turn it up?” and someone says, “Sure.” That’s fine. That is how fitting works. It just means the weights belong to the dataset, not to you.&lt;&#x2F;p&gt;
&lt;p&gt;I say this as a person whose entire job is explaining weighted averages to strangers, and who could not tell you his own lymphocyte percentage under oath.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-these-calculators-cannot-settle&quot;&gt;What these calculators cannot settle&lt;&#x2F;h2&gt;
&lt;p&gt;The facts here do not settle how well blood-based calculators agree with each other across labs, and nobody has published a shared reference standard for them in what I have. The closest evidence is from the epigenetic side, which is not encouraging. The Scientist reported that as of 2026 the epigenetic clock field “has not yet achieved reference standards,” so it is not possible to send one sample to different validated labs and get the same readout. Elysium, which sells one, says a single epigenetic sample processed repeatedly can give estimates that vary by up to nine years. Nine years is a modest amount of drift for one tube of spit or blood.&lt;&#x2F;p&gt;
&lt;p&gt;There is also the disclosure question. Index is open about its relationship with PhenoAge. Most companies selling biological age kits, per The Scientist, do not detail which clock they use, which is a discreet way to run a relationship.&lt;&#x2F;p&gt;
&lt;p&gt;Whether moving the number means much is also unsettled. In the CALERIE trial, long-term caloric restriction did not significantly change biological age estimates measured by PhenoAge or GrimAge. And a commentary in the AMA Journal of Ethics argues that these tests “cannot tell us how old a person ‘really’ is,” calls biological age a fiction, and warns that the results can make age-related discrimination look justifiable. Etc.&lt;&#x2F;p&gt;
&lt;p&gt;One small detail sits underneath all of it. The seventh input to GeroAge, after the six blood markers, is your chronological age, which is the only number on the form that no lab has to measure.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>Biological Age Tests: What They Measure and What They Don&#x27;t</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/biological-age-test/"/>
        <id>https://bioclocktest.com/biological-age-test/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/biological-age-test/">&lt;p&gt;I have had a birthday every year of my life, and lately I have come to admire it as a measuring instrument: it has never once come back nine years off. A biological age test tries to give you a second number, an age read off your body instead of the calendar, by measuring something in a sample (usually blood) and feeding it into a formula trained to predict age, disease or death. The result arrives in years, which makes it look like the same kind of thing as a birthday.&lt;&#x2F;p&gt;
&lt;p&gt;It is not quite. There are four main kinds: methylation tests, which read chemical tags on your DNA; blood chemistry tests, which reweight ordinary lab values like glucose and albumin; proteomic tests, which measure thousands of blood proteins; and telomere tests, which measure the ends of chromosomes. What none of them can do, the AMA Journal of Ethics argued in December 2025, is tell you how old a person “really” is, and as of 2026 the field has no reference standards, so the same sample can come back with different ages from different labs.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-four-kinds-of-biological-age-test-and-the-one-recipe-behind-them&quot;&gt;The four kinds of biological age test and the one recipe behind them&lt;&#x2F;h2&gt;
&lt;p&gt;Every one of these is built the same basic way. You take a big group of people about whom you know a lot (their ages, their lab results, later their diagnoses and deaths), measure some large pile of stuff in their bodies, and fit a formula that turns the pile into a number. Then you apply the formula to a new person, who gets a number in years.&lt;&#x2F;p&gt;
&lt;p&gt;The differences are in what goes into the pile and what the formula was trained to predict. Consumer tests, per one explainer, measure metabolic markers, analyze DNA directly, or analyze proteins in blood. Telomeres are the older dish still printed on some menus.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;methylation-clocks-read-chemical-tags-on-blood-dna&quot;&gt;Methylation clocks read chemical tags on blood DNA&lt;&#x2F;h2&gt;
&lt;p&gt;Your DNA carries small chemical tags at particular spots, and the pattern drifts with age in ways regular enough to fit. The Horvath clock, published in 2013, was the landmark: a multi-tissue clock that predicts age across blood, brain, liver and various cell types. The Hannum clock uses 71 CpG sites (the specific spots where tags get read) and performs well in blood specifically. A 2018 paper claimed methylation clocks beat every other omics- or clinical-data-based aging biomarker on accuracy.&lt;&#x2F;p&gt;
&lt;p&gt;How accurate is accurate? One recent method, evaluated on 3,707 blood methylation samples, hit a mean absolute error of 3.149 years, meaning its guesses at plain chronological age missed by about three years on average. That is a good result for the field.&lt;&#x2F;p&gt;
&lt;p&gt;It is also three years.&lt;&#x2F;p&gt;
&lt;p&gt;The first clocks were trained to predict how old you are, which is a slightly odd product, since you already know (it is printed on your driver’s license, for free). Later clocks were trained on things the calendar does not know. Morgan Levine identified methylation sites that tracked nine clinical chemistry biomarkers, using data from the National Health and Nutrition Examination Survey. Belsky and colleagues went further in 2020 with what eLife described as “a single-time-point blood test measure of a person’s pace of biological aging.” Its successor, DunedinPACE, was trained only on probes that passed a test-retest reliability screen, which in some sense is the most important design choice in this whole article.&lt;&#x2F;p&gt;
&lt;p&gt;The training target decides what moves the number. In the Add Health study of younger U.S. adults, sociodemographic and lifestyle factors were more often associated with clocks trained on current or dynamic phenotypes (PhenoAge, GrimAge, DunedinPACE) than with clocks trained on age alone, like Horvath. A clock built to predict your birthday is, reasonably, not that interested in your habits. Also, almost all commercial clocks are trained on blood, though skin age can be read from tape stuck to your face.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;blood-chemistry-clocks-reweight-your-annual-physical&quot;&gt;Blood chemistry clocks reweight your annual physical&lt;&#x2F;h2&gt;
&lt;p&gt;The second family skips DNA and uses labs a doctor already orders. Typical markers include glucose, LDL cholesterol, triglycerides, testosterone, SHBG, DHEAS, the liver enzyme GGT and white blood cells; one explainer series adds CRP, albumin, alkaline phosphatase, creatinine, red cell distribution width, mean corpuscular volume and lymphocytes. None of this is exotic. Most people have had these drawn while thinking about something else entirely (usually the needle).&lt;&#x2F;p&gt;
&lt;p&gt;So what you are buying is the weighting. InsideTracker’s calculation, InnerAge 2.0, includes glucose and GGT and gives glucose more weight because glucose correlates more strongly with age in InsideTracker’s own dataset. That’s fine; that is what fitting a model means. It also means your answer depends on whose dataset did the fitting.&lt;&#x2F;p&gt;
&lt;p&gt;The weights are the product.&lt;&#x2F;p&gt;
&lt;p&gt;And you need fewer inputs than you might think. GeroAge uses six blood markers plus chronological age and agrees with PhenoAge at R=0.99, with similar power to predict survival in NHANES data. A hematology clock built in mice found platelets were the strongest predictor of biological age (the mice were not told their results).&lt;&#x2F;p&gt;
&lt;h2 id=&quot;proteins-telomeres-and-the-organ-ages&quot;&gt;Proteins, telomeres and the organ ages&lt;&#x2F;h2&gt;
&lt;p&gt;The Proteomic Aging Clock, built in the UK Biobank Pharma Proteomics Project from 53,021 people and 2,923 proteins, showed stronger age-adjusted associations with all-cause mortality and 18 incident diseases than PhenoAge, BioAge or leukocyte telomere length. The proteins that drive its deviations cluster in immunoinflammatory responses, cellular senescence, extracellular matrix remodeling and vascular biology. Another proteomic clock tracked telomere length, a frailty index and reaction time.&lt;&#x2F;p&gt;
&lt;p&gt;That is roughly all these facts say about telomeres: they appear here mainly as the thing the protein clock beat. The site’s telomere article handles the rest; whether they belong in this lineup at all is not settled here.&lt;&#x2F;p&gt;
&lt;p&gt;Then there are organ ages. TruDiagnostic’s SymphonyAge reports the biological age of 11 organ systems, and its TruAge product reports 64 longevity biomarkers. Elysium Health’s Index reports nine body systems, counting inflammation and hormones as systems. Your body has nine systems or eleven.&lt;&#x2F;p&gt;
&lt;p&gt;Depends who is counting.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-the-same-blood-can-come-back-with-different-ages&quot;&gt;Why the same blood can come back with different ages&lt;&#x2F;h2&gt;
&lt;p&gt;Elysium’s own science page says a single sample processed repeatedly can produce estimates varying up to nine years, and that Index agrees with itself across technical replicates within zero to 1.5 years. Picture a bathroom scale that is accurate on average and wanders by nine years per step. The Scientist adds that “most companies offering biological age test kits do not detail which epigenetic clock they use.”&lt;&#x2F;p&gt;
&lt;p&gt;Everyone here is acting sensibly from where they stand. The company spent real money on its clock and its lab processing, and naming them hands a roadmap to competitors. The academics built clocks to compare groups, where noise averages out. The customer paid for a number, not a methods section. So somewhere a product manager says “let’s call it our proprietary algorithm and get back to the packaging,” and that is a reasonable thing to say.&lt;&#x2F;p&gt;
&lt;p&gt;Here is a scenario I am inventing. You split one tube of blood ten ways, mail it to ten labs, and get ten ages spread over nine years. You report the youngest. A startup offers to do the splitting for you. Labs notice that the customers who come back are the ones who got young numbers, so labs quietly compete on youth. Someone launches a fund that holds all ten labs and reports the minimum. I feel like this would be the most effective anti-aging intervention ever sold, and it involves no biology.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;three-ways-to-read-a-biological-age-result&quot;&gt;Three ways to read a biological age result&lt;&#x2F;h2&gt;
&lt;p&gt;There are, roughly, three ways to read this.&lt;&#x2F;p&gt;
&lt;p&gt;The first is that a biological age is a risk score reported in years. The ethics commentary itself says these tests have been shown to be “fairly reliable indicators” of a person’s risk of certain age-related diseases. Years are a friendly unit for risk, if a slightly dumb one, and on this reading the unit is harmless.&lt;&#x2F;p&gt;
&lt;p&gt;The second is that it is a marketing channel. Tony Wyss-Coray called the consumer market “a bit like the wild West,” noting that many companies sell tests coupled with products never tested in humans in a clinical trial. A test on its own is a modest business. A test that hooks up with a product line is a sexier arrangement, because the follow-up test is how you find out whether the relationship is working, and I guess you keep paying to find out.&lt;&#x2F;p&gt;
&lt;p&gt;The third is the uncomfortable one. The same commentary calls biological age a fiction, warns results can cause psychological harm and make age-related discrimination appear justifiable, and asks companies to be more transparent about reliability and limits. Look, the obvious fix for a noisy number is a better number. But a better number looks more official, and an official number about how old someone really is becomes exactly the sort of thing a person can be judged by. Accuracy does not solve that.&lt;&#x2F;p&gt;
&lt;p&gt;It makes it worse.&lt;&#x2F;p&gt;
&lt;p&gt;My birthday will come around again next year, off by zero, predicting nothing about my kidneys. It agrees with itself because it measures nothing interesting. The interesting numbers are the ones that disagree.&lt;&#x2F;p&gt;
&lt;p&gt;So far nobody sells a number that does both.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>Biological Age vs Chronological Age: What a Gap Means</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/biological-age-vs-chronological-age/"/>
        <id>https://bioclocktest.com/biological-age-vs-chronological-age/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/biological-age-vs-chronological-age/">&lt;p&gt;The University of California, San Francisco keeps a web page about epigenetic clocks, and on it is one sentence that does most of the work for this entire industry. Here it is: “Epigenetic Age Acceleration (AgeAccel) is defined as the residual from regressing DNAm Age on chronological age, where a positive value indicates epigenetic age is greater than expected.” That is the official definition of being older than your age. It rewards slow reading.&lt;&#x2F;p&gt;
&lt;p&gt;The short answer for anyone who searched biological age vs chronological age: chronological age is how long you have been alive, and biological age is a number a model produces from your blood chemistry or your DNA methylation, built so that across a big group of people it tracks chronological age closely. The gap between the two is the part people care about. A positive gap means the model placed you older than people your age usually come out, and a negative gap means younger. What that gap says about your actual body is not settled, and for many tests the gap you get is smaller than the test’s own noise.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;biological-age-vs-chronological-age-read-clause-by-clause&quot;&gt;Biological age vs chronological age, read clause by clause&lt;&#x2F;h2&gt;
&lt;p&gt;Start with “Epigenetic Age Acceleration.” This sounds like a speedometer. It is not. In the UCSF definition it is a single number from a single blood draw, which means it measures position, not speed. A thing measured once cannot accelerate. It can only be somewhere.&lt;&#x2F;p&gt;
&lt;p&gt;Next, “DNAm Age.” This is the output of a clock that reads methylation, the chemical tags on DNA. The Horvath clock uses 353 sites (CpG loci, if you want the vocabulary) and the Hannum clock uses 71. The sites were chosen, per UCSF, by a statistical method “aimed at producing the most accurate estimate of chronological age.” Hold onto that phrase. It matters later.&lt;&#x2F;p&gt;
&lt;p&gt;Then “the residual from regressing DNAm Age on chronological age.” Translated into one flat sentence: take a lot of people, plot the clock’s age against their real age, draw the best straight line through the cloud, and measure how far above or below the line your dot sits. That distance is your acceleration, and a consumer testing business is stacked on top of it (which is a lot of weight for one residual to carry).&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-greater-than-expected-is-measured-against&quot;&gt;What “greater than expected” is measured against&lt;&#x2F;h2&gt;
&lt;p&gt;Now some formality, because the point deserves it. For the avoidance of doubt, a person’s age acceleration is not their clock age minus their birthday age. It is their clock age minus the clock age the fitted line assigns to people of their birthday age. These two quantities coincide only where the line happens to pass exactly through the diagonal, which it need not do.&lt;&#x2F;p&gt;
&lt;p&gt;Here is an invented example, clearly made up. Say you are 50 and the clock says 54. In a sample where the line puts the typical 50-year-old at 50, your residual is +4. In a sample where the line puts the typical 50-year-old at 53 (because of the lab, the batch, the population, or whatever), your residual is +1. Same blood. Same 54. Different neighbors.&lt;&#x2F;p&gt;
&lt;p&gt;So “expected” means expected by a line, and the line was drawn through some particular group of people. It is sort of like a golf handicap, except nobody tells you which course you played. And between us, if we got to pick the comparison group, we would both pick the one that makes us look young.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-the-first-clocks-treat-their-own-mistakes-as-the-finding&quot;&gt;Why the first clocks treat their own mistakes as the finding&lt;&#x2F;h2&gt;
&lt;p&gt;This is the part I find delightful, which tells you something about my weekends.&lt;&#x2F;p&gt;
&lt;p&gt;The first-generation clocks, Horvath and Hannum, were built to predict chronological age as accurately as possible. The Horvath clock gets within a mean absolute error of about 3.6 years across tissues, and these clocks correlate with chronological age at r above 0.90 in adults aged 20 to 100. That is very good at telling you something you already knew from your driver’s license (which is cheaper, and rarely off by 3.6 years).&lt;&#x2F;p&gt;
&lt;p&gt;Then the residual, the part the model got wrong, gets renamed “acceleration” and treated as the signal. If the clock were perfect, every residual would be zero and nobody would be older or younger than anybody. The clock’s usefulness lives entirely in its failure to do the job it was trained for. I mean, it’s a strange arrangement, but it is the arrangement.&lt;&#x2F;p&gt;
&lt;p&gt;To be fair, the field noticed. The second-generation clocks, PhenoAge (published in 2018) and GrimAge, changed the training target from chronological age to clinical measurements, disease incidence and death. DunedinPACE is sometimes called third generation because it was built from data tracking physiological decline over time in people born the same year. And it shows up in results: in the Add Health study of younger U.S. adults, sociodemographic and lifestyle factors were more often associated with aging in PhenoAge, GrimAge and DunedinPACE than in Horvath. The clocks built to find age found age. The clocks built to find health found more of the other stuff.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-being-older-can-actually-be-picking-up&quot;&gt;What being “older” can actually be picking up&lt;&#x2F;h2&gt;
&lt;p&gt;A methylation clock usually reads blood, and blood is a mix of cell types. Change the mix and you change the reading, whether or not anything about your aging changed. UCSF lists a version called intrinsic acceleration that also regresses out blood cell counts, and an extrinsic version that upweights cell composition on purpose to capture aging of the immune system. Same tube, two definitions, two answers.&lt;&#x2F;p&gt;
&lt;p&gt;This matters for real findings. First-generation clocks generally put people living with HIV about 5 to 7 years older than HIV-negative people of the same age, though a review notes this may be confounded by HIV-related changes in immune cell composition. The line between “your cells are aging” and “your tube contains a different crowd of cells” is, let’s say, blurry, and everyone involved is consenting adults about it.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-big-a-gap-has-to-be-before-it-means-anything&quot;&gt;How big a gap has to be before it means anything&lt;&#x2F;h2&gt;
&lt;p&gt;The published effects are often small. In one cohort, having four or more adverse childhood experiences was associated with GrimAge acceleration of 0.71 years (95% CI: 0.42 to 1.00) at a mean age of about 40, after adjusting for demographics, behavior and income. That is a real population-level finding. It is also less than a year.&lt;&#x2F;p&gt;
&lt;p&gt;Compare that with measurement noise. Elysium Health, which sells an aging test, says on its own science page that “a single epigenetic age sample processed repeatedly can result in biological age estimates that vary up to nine years.” The Scientist reports that as of 2026 the field has no reference standards, so you cannot send one sample to several validated labs and get the same number back. Blood-chemistry clocks can produce spreads of 20 years younger to 20 years older than chronological age. I don’t know how much of that spread is biology and how much is net.&lt;&#x2F;p&gt;
&lt;p&gt;Now a scenario, entirely made up. A lab tech runs your sample and gets 46. “Hey, the plate looked weird, we reran it,” she says, “you’re 53 now.” Run it seven more times and you have nine ages. A sufficiently motivated company could report the median, or the youngest, or a weighted average of the youngest three, and call it a proprietary noise-reduction algorithm. Then a competitor reports the youngest two. By the end of the year everyone is 31, lol.&lt;&#x2F;p&gt;
&lt;p&gt;For instance, the AMA Journal of Ethics ran a commentary arguing these tests cannot tell us how old a person “really” is, and calling biological age a “fiction.” That is harsher than the field would put it. But the point is that the gap is a residual from a line, the line belongs to a sample, and the sample belongs to a lab.&lt;&#x2F;p&gt;
&lt;p&gt;In a Russian plasma protein study, the order in which samples were run through the mass spectrometer predicted participants’ age with a mean absolute error of 3.98 years.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>Biological Age, Defined Carefully</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/biological-age/"/>
        <id>https://bioclocktest.com/biological-age/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/biological-age/">&lt;p&gt;Today the epigenetic clock field “has not yet achieved reference standards,” according to The Scientist, which means it is not possible to send one sample to different validated labs and get the same age back. That is a weird property for a measurement to have. A ruler is allowed to be wrong. It is just expected to be wrong the same way twice.&lt;&#x2F;p&gt;
&lt;p&gt;Here is the plain answer. Biological age is an estimate, stated in years, of how old your body seems based on measurements of it (DNA methylation, blood chemistry, proteins, etc.), as opposed to how many years have passed since you were born. There is no single agreed definition. Some clocks are built to guess your calendar age, some to guess your risk of getting sick or dying, and some to guess how fast you are changing, and they give different numbers. A December 2025 commentary in the AMA Journal of Ethics goes further and calls the concept a “fiction.”&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-chronological-age-is-a-weak-proxy&quot;&gt;Why chronological age is a weak proxy&lt;&#x2F;h2&gt;
&lt;p&gt;The case for the whole enterprise starts with the birth certificate. It is an excellent document with one flaw: it gives the same number to everyone born in the same year. A review of proteomic aging clocks states the motivation carefully, saying biological age “may provide a more informative metric of individual physiological function and life expectancy compared to chronological age.” (The “may” is doing honest work there.) Two 50-year-olds can differ in basically every way a doctor cares about, and the calendar cannot tell them apart.&lt;&#x2F;p&gt;
&lt;p&gt;How far apart do people get? One 25-biomarker blood clock produced estimates ranging from 20 years younger to 20 years older than people’s chronological age. That is a 40-year spread among people who all checked the same box on a form. A study in a Pakistani population found epigenetic age prediction got less accurate as people got older, which the authors read as environment and lifestyle mattering more later in life. Baseball scouts have always known a version of this (a 29-year-old shortstop with old legs is, for scouting purposes, not 29), and the clocks are an attempt to turn the scout’s hunch into a lab report.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-dumb-version-a-clock-that-guesses-your-birthday&quot;&gt;The dumb version: a clock that guesses your birthday&lt;&#x2F;h2&gt;
&lt;p&gt;The simple way to build a biological age goes like this. Take a lot of people whose ages you already know, measure something about them, and train a model to guess their age from the measurement. The Horvath clock was trained to predict chronological age alone. The Hannum clock uses 71 CpG loci and performs well in blood. A deep-learning model built from NHANES data, covering 24,985 people surveyed from 1999 to 2010, predicted chronological age with a mean absolute error of 2.98 years.&lt;&#x2F;p&gt;
&lt;p&gt;This is a strange product, if you think about it. A clock that guessed everyone’s age perfectly would be useless, because the answer is already in a drawer at home. The value is entirely in the misses. So the field formalizes the misses: in the definition used by UCSF, epigenetic age acceleration “is defined as the residual from regressing DNAm Age on chronological age,” and a positive value means your epigenetic age is higher than expected.&lt;&#x2F;p&gt;
&lt;p&gt;It works like splitting a dinner bill evenly and then checking who ordered the lobster. Everyone your age is charged the table average. Your acceleration is how far your own order sits from that average. It’s fine as accounting. On the other hand, a model trained to hit everyone’s age is, by construction, trying not to miss, and in the Add Health study of younger U.S. adults, sociodemographic and lifestyle factors were less often associated with clocks trained on chronological age alone (like Horvath) than with clocks trained on other targets.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-sophisticated-version-a-clock-that-guesses-your-health&quot;&gt;The sophisticated version: a clock that guesses your health&lt;&#x2F;h2&gt;
&lt;p&gt;So later clocks aimed at something else. Morgan Levine built PhenoAge by finding DNA methylation sites correlated with nine clinical chemistry biomarkers and chronological age, again using NHANES. GrimAge and DunedinPACE were also trained on current or dynamic phenotypes rather than birthdays. A proteomic aging clock called PAC “was developed as a proxy of biological age targeting all-cause mortality risk,” and the proteins tied to it were enriched in things like immunoinflammatory responses, cellular senescence and vascular biology. DunedinPACE goes furthest and does not report an age at all; it is described as monitoring the rate of aging rather than predicting a static biological age.&lt;&#x2F;p&gt;
&lt;p&gt;This is more ambitious. The question is no longer what your passport says. It is how your body is doing, measured by what tends to happen to bodies like it. And in Add Health, lifestyle and social factors showed up more often in these clocks. That is not nothing.&lt;&#x2F;p&gt;
&lt;p&gt;Then comes the product meeting. “Great, it tracks mortality risk,” someone says. “Can it say 47?” Mostly it can, and mostly it does. TruDiagnostic describes its OMICmAge clock as one that “quantifies biological age against chronological age,” which is the dinner bill again, with a better menu.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;where-the-two-versions-of-biological-age-end-up&quot;&gt;Where the two versions of biological age end up&lt;&#x2F;h2&gt;
&lt;p&gt;Here is the part I find delightful. Researchers built GeroAge, a simplified clock needing only six blood markers (albumin, lnCRP, lymphocyte percentage, creatinine kinase, blood glucose and white blood cell count) plus chronological age. It showed R=0.99 agreement with PhenoAge and similar predictive power for overall survival in NHANES.&lt;&#x2F;p&gt;
&lt;p&gt;Sit with 0.99 for a second. That is the correlation you expect between a document and a photocopy of the document. Six routine blood numbers and a birthday reproduce, almost exactly, the more elaborate measure. And chronological age is one of the inputs. Basically the calendar was in the recipe the whole time.&lt;&#x2F;p&gt;
&lt;p&gt;So the long road runs like this. Start with the birth certificate, add blood or methylation, train on illness and death instead of birthdays, translate the result back into years, and subtract the birth certificate to see what is left. The dumb road was: guess the birth certificate and look at what is left. Both arrive at a number in years whose meaning is “how far from the calendar.” The sophisticated one has better reasons for its residual. It is still a residual.&lt;&#x2F;p&gt;
&lt;p&gt;The field keeps producing new contenders anyway. EpiCAge, which combines epigenetic and phenotypic data, consistently outperformed existing clocks across eight internal and four external cancer cohorts. iTARGET is reported to outperform traditional epigenetic clocks too. (Every methods section I read outperforms every other one, which is also how my newsletter drafts look to me before an editor gets to them.) Whether any of them has settled what biological age is, rather than how to predict it better, is not settled.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-a-biological-age-on-a-consumer-report-is&quot;&gt;What a biological age on a consumer report is&lt;&#x2F;h2&gt;
&lt;p&gt;On the consumer side, the number arrives in an app. No direct-to-consumer biological age clock is approved by the FDA as a medical diagnostic; they are sold as wellness or informational products. Most companies selling kits do not say which epigenetic clock they use (presumably to keep the packaging uncluttered). Elysium Health, which does say (its Index uses DNAm PhenoAge), also says that a single sample processed repeatedly can produce estimates that vary by up to nine years. That is a fairly wide spread. Tony Wyss-Coray of Stanford described the market as “a bit like the wild West.”&lt;&#x2F;p&gt;
&lt;p&gt;The reports have also started subdividing. Index gives ages for nine body systems, including brain, liver, kidney and heart. TruDiagnostic’s SymphonyAge measures 11 organ systems. I suppose this follows the logic all the way down. Picture a person (I am inventing her) whose report says her liver is 34, her kidneys are 51 and her immune system is 42. Her liver wants a quiet birthday dinner and her kidneys want a retirement party. She sends the same sample in again, and her kidneys come back 43, and now they would like to un-retire. If she mails it to a second lab, she gets a third family.&lt;&#x2F;p&gt;
&lt;p&gt;None of this makes the clocks worthless as research tools; they correlate with real things, and the trials and disease studies are covered elsewhere on this site. The AMA Journal of Ethics worry is about what happens outside research: that a result like this “cannot tell us how old a person ‘really’ is,” and that it can make age-related discrimination look justifiable, because a number with a decimal point looks like a fact. Meanwhile the birth certificate keeps issuing the same answer every time it is consulted, which was always its only flaw, and which, compared with the alternatives, is starting to look like a feature.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>Bryan Johnson&#x27;s Biological Age and What One Person Can Measure</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/bryan-johnson-biological-age/"/>
        <id>https://bioclocktest.com/bryan-johnson-biological-age/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/bryan-johnson-biological-age/">&lt;p&gt;If you are Bryan Johnson, or anyone else who has decided to measure their own aging as often and as thoroughly as money allows, you end up with a great many numbers about exactly one body. People who type Bryan Johnson biological age into a search box are usually looking for one of those numbers: a single figure that says how old he “really” is. The sources behind this article do not report his results, so there is no figure here to hand over, and I am not going to invent one. What the sources do settle is how far any one person’s biological age reading can be trusted. The answer is less far than the decimal point suggests.&lt;&#x2F;p&gt;
&lt;p&gt;The core problem is that a biological age test is a noisy measurement, and one person is a sample size of one. Elysium Health, which sells such a test, says on its own science page that “a single epigenetic age sample processed repeatedly can result in biological age estimates that vary up to nine years.” Separately, The Scientist reports that as of 2026 the field has no reference standards, so “it is not possible to send a sample to different validated labs and obtain the same biological age readout.” Any self-tracker’s best number sits inside that fog, however famous the self-tracker is.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-dumb-version-take-one-test-read-one-number&quot;&gt;The dumb version: take one test, read one number&lt;&#x2F;h2&gt;
&lt;p&gt;Here is the simple way to learn your biological age. You buy a kit, you provide some blood or cheek cells, a lab reads chemical tags on your DNA, and an algorithm turns those tags into an age. You read the age. You feel good or bad about it. Done.&lt;&#x2F;p&gt;
&lt;p&gt;The problems with this version are the obvious ones. The Scientist notes that most companies selling these kits “do not detail which epigenetic clock they use,” so you may not know which of several quite different algorithms produced your number. (I assume this is because the clock names are long and the packaging is small.) None of these consumer clocks is approved by the Food and Drug Administration as a medical diagnostic tool; they are sold as wellness or informational products. And the same tube, run twice, can come back years apart.&lt;&#x2F;p&gt;
&lt;p&gt;Picture the customer service call. “Hey, so, it’s the same sample. We just ran it again. You are now 41 instead of 36. That’s within the expected range.” The person on the phone is doing their job correctly. The job is just sort of strange.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-your-best-reading-is-probably-too-good&quot;&gt;Why your best reading is probably too good&lt;&#x2F;h2&gt;
&lt;p&gt;There is a second, quieter problem, which is regression to the mean. It is a statistics idea, and it works like this. Every measurement is the true value plus some random error. If you take one reading and it comes out unusually high, part of the reason is probably that the error happened to be big and positive that day. Measure again and the error will presumably be more ordinary, so the second reading will tend to land closer to the middle. Nothing changed in you. The dice just rolled differently.&lt;&#x2F;p&gt;
&lt;p&gt;The dull everyday version is a bathroom scale. You step on, it says a number you dislike, you step off, step on again, and it says something a little better. If you keep doing this and write down only the reading you like, your record will show steady weight loss. It will show this whether or not you lost any weight.&lt;&#x2F;p&gt;
&lt;p&gt;Now attach a behavior to it. A person gets a bad biological age result, which is exactly the moment people start new regimens. They retest a few months later and the number improves. Some of that improvement may be real. Some of it is the bad first reading drifting back toward the person’s actual level, which it would have done anyway. With one person and no comparison group, the two are, in some sense, impossible to separate.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-sophisticated-version-many-clocks-many-draws-one-person&quot;&gt;The sophisticated version: many clocks, many draws, one person&lt;&#x2F;h2&gt;
&lt;p&gt;So the serious self-tracker does the obvious upgrade. Don’t take one test; take lots. Use several clocks, repeat the draws, average them, and track trends instead of points. This is what a careful person with a big budget would do, and to be fair it fixes real things.&lt;&#x2F;p&gt;
&lt;p&gt;It also introduces new complications, because the clocks are not measuring the same thing. Some, like Horvath’s original clock, were trained to predict chronological age. Others, like PhenoAge and GrimAge, were trained on health outcomes, and DunedinPACE is a pace measure (a speed, not an age) that its eLife developers describe as “a single-time-point blood test measure of a person’s pace of biological aging.” In the Add Health study of younger American adults, lifestyle and social factors were more often associated with the outcome-trained clocks than with Horvath. A hormone study found the cortisol&#x2F;DHEAS ratio correlated with some clocks (Hannum, Horvath Skin &amp;amp; Blood, PhenoAge) and not others (Horvath, GrimAge). So when your numbers disagree, that is not necessarily a malfunction. It may be the design.&lt;&#x2F;p&gt;
&lt;p&gt;Then there are the adjustments. A UCSF description puts one of them this way: “Intrinsic Epigenetic Age Acceleration (IEAA) is derived from the Horvath clock by regressing epigenetic age on chronological age and blood cell count measures.” In flat words: they take your clock age, subtract what your real age and your mix of blood cells would have predicted, and call the leftover your acceleration. That leftover is a residual. Residuals are where noise goes to live.&lt;&#x2F;p&gt;
&lt;p&gt;And interventions do not reliably move every clock. In the CALERIE trial, years of calorie restriction did not significantly change PhenoAge or GrimAge estimates (the site’s CALERIE article covers which measure did move). That was a randomized trial with many people. A single person trying to detect something smaller with fewer measurements is in a harder spot.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-one-person-can-measure-about-bryan-johnson-s-biological-age&quot;&gt;What one person can measure about Bryan Johnson’s biological age&lt;&#x2F;h2&gt;
&lt;p&gt;Which brings the long road back to the dumb version’s front door. With many clocks and many draws, a self-tracker learns a few useful things: how much their own readings bounce around, which clocks agree with each other for them, and whether a change is large relative to that bounce. Those are real gains (if not quite the gains the brochure had in mind). A change bigger than the noise, sustained across repeated tests from the same lab and the same clock, is more believable than one lucky Tuesday.&lt;&#x2F;p&gt;
&lt;p&gt;What they cannot learn is cause. A dedicated tracker usually changes dozens of things at once (diet, sleep, exercise, supplements, presumably the brand of water) and has no untreated twin to compare against. Even a clean downward trend is a trend in one person, with regression to the mean and lab drift baked in.&lt;&#x2F;p&gt;
&lt;p&gt;Look, I write about this for a living, which mostly means reading footnotes about pipetting variability and calling it journalism. But the footnotes matter here. You and I both know that a number printed to one decimal place feels more true than a number with a nine-year wobble attached.&lt;&#x2F;p&gt;
&lt;p&gt;Here is a made-up scenario, to show where the logic goes. A person tests weekly. Because one sample can swing by years, some week will come back spectacular: say, biologically 29 at chronological 46. They repeat whatever they did that week. Next week is worse, so they assume they slipped, and adjust. Over a year they have optimized their life around the lab’s random error, like Bill Murray in Groundhog Day, except the day never gets better because the day was never the problem. Eventually they are eating the same lunch at the same minute because it coincided with a good draw. The clock has trained them. That is sort of thing a model is supposed to do to data, not people.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;where-the-long-road-ends&quot;&gt;Where the long road ends&lt;&#x2F;h2&gt;
&lt;p&gt;The research view is not that these tests are worthless. The AMA Journal of Ethics commentary says epigenetic clocks have been shown to be “fairly reliable indicators of a person’s risk of developing certain age-related diseases.” That is a statement about risk across groups, which is what the clocks were built and validated on. The same commentary argues the tests cannot tell anyone how old they “really” are, and calls biological age a “fiction.”&lt;&#x2F;p&gt;
&lt;p&gt;Those two statements fit together fine. A clock can sort a thousand people by risk reasonably well while telling you very little about where exactly one of them sits, the same way a rough poll can call an election and still misreport your neighbor. The sophisticated self-tracker, with every clock and every repeat draw, ends up with a better estimate of their own noise and a cautious trend line. The dumb version ends up with one number and a feeling. Whether Bryan Johnson’s numbers are good, middling or unusually young is not something the evidence here can settle, and it is not obvious that more testing on him alone ever could.&lt;&#x2F;p&gt;
&lt;p&gt;One person is still one person, however many times you count him.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>The CALERIE Trial: Calorie Restriction Slowed a Clock</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/calerie-trial/"/>
        <id>https://bioclocktest.com/calerie-trial/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/calerie-trial/">&lt;p&gt;A group of researchers took 220 healthy adults, told roughly two-thirds of them to eat a quarter less food for two years, and let the rest carry on as usual. Later, other researchers ran the participants’ stored blood through three different epigenetic aging clocks. One clock said the eat-less group was aging a little slower. The other two did not notice anything.&lt;&#x2F;p&gt;
&lt;p&gt;That is basically the result of the CALERIE trial, whose full name is the Comprehensive Assessment of Long-term Effects of Reducing Intake of Energy (someone worked hard on that acronym). Adults without obesity were randomized to 25% caloric restriction or an “ad libitum” control diet for 2 years. In a post-hoc analysis, restriction slowed DunedinPACE, a clock that estimates the speed of aging, by approximately 3%. It did not produce statistically significant changes in two other clocks, PhenoAge and GrimAge. The authors call the effect sizes small. Whether a 3% slower clock means anyone lives longer is not settled.&lt;&#x2F;p&gt;
&lt;p&gt;The published analysis compresses all of this into one sentence: “In the CALERIE randomized controlled trial, 25% caloric restriction over 2 years slowed DunedinPACE by approximately 3%, but did not produce statistically significant changes in PhenoAge or GrimAge.” It is a careful sentence. If you read it slowly, the way you would read a contract you are about to sign, it turns out to contain most of what is currently known about biological age tests, and most of what isn’t.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-the-calerie-trial-actually-did&quot;&gt;What the CALERIE trial actually did&lt;&#x2F;h2&gt;
&lt;p&gt;Start with “randomized controlled trial.” This is the expensive clause. Most of what anyone knows about biological age comes from observing people: smokers have older-looking blood, richer people have younger-looking blood, etc. Observation cannot tell you whether the habit moved the clock or whether the kind of person who has the habit was always going to have that clock. Randomization fixes this by letting a coin decide (coins have no opinions about kale).&lt;&#x2F;p&gt;
&lt;p&gt;The coin here was weighted. Participants were assigned to restriction or control in a 2:1 ratio. Picture a dinner party where the host flips a coin for each guest, except the coin is rigged so two out of every three guests get a plate one quarter smaller, and nobody gets to swap. The guests were healthy, nonobese men and premenopausal women aged 21 to 50, with BMIs between 22.0 and 27.9. These were not people with a weight problem to solve. They were ordinary-sized people asked to become slightly smaller ordinary-sized people.&lt;&#x2F;p&gt;
&lt;p&gt;The trial is registered on ClinicalTrials.gov as NCT00427193, which matters only in the sense that it existed, on paper, before anyone looked at a methylation clock. That is the other half of why randomized trials are worth their trouble.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-25-caloric-restriction-over-2-years-covers&quot;&gt;What “25% caloric restriction over 2 years” covers&lt;&#x2F;h2&gt;
&lt;p&gt;“25%” is what people were assigned. It is not necessarily what everyone did. A separate analysis of the same trial, looking at telomeres, measured something it calls caloric restriction “dose” and found it varied enough to be worth analyzing, which is a polite way of saying that people asked to eat 25% less ate various amounts less.&lt;&#x2F;p&gt;
&lt;p&gt;“Over 2 years” is the understated clause. Two years of eating three-quarters of what you would otherwise eat is, in the technical sense, a commitment. It is several hundred breakfasts (plus the lunches, which the sentence also skips). The sentence spends four words on it.&lt;&#x2F;p&gt;
&lt;p&gt;And then the blood went into freezers. The CALERIE Genomic Data Resource was later generated from stored samples from 218 participants (two of the original 220 are not accounted for in that count, and the facts available here do not say why). It includes genotypes and three timepoints of DNA methylation, mRNA and small RNA data, from blood, skeletal muscle and fat. The total is 2,327 samples.&lt;&#x2F;p&gt;
&lt;p&gt;It is worth sitting with 2,327 for a second. That is not a round number. Nobody planned 2,327. It is what you get when you take a couple hundred people, cut small pieces out of them at three points over two years, and keep every piece that survived the freezer. Every one of those samples is a person on a smaller plate.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-the-dunedinpace-number-is-the-one-that-moved&quot;&gt;Why the DunedinPACE number is the one that moved&lt;&#x2F;h2&gt;
&lt;p&gt;“Slowed DunedinPACE” is a weird verb with a weird object. Most clocks estimate an age: your blood looks 47. DunedinPACE was trained differently, on longitudinal change rather than on predicting age at one moment. It is meant to estimate a rate, how fast a body is deteriorating right now, which is why the sentence says the clock “slowed” rather than that anyone got younger. The site’s article on DunedinPACE covers how it is built.&lt;&#x2F;p&gt;
&lt;p&gt;This matters for which clock you would expect to move in a two-year trial. The Moffitt &amp;amp; Caspi lab, which developed it, describes DunedinPACE as “the most sensitive epigenetic clock to intervention,” slowing most in trials and responding most consistently across studies. In trial-design terms that makes it the sexy clock: the one most likely to respond to an intervention’s advances. On the other hand, a clock that reacts to everything is also a clock you have to be careful about falling for.&lt;&#x2F;p&gt;
&lt;p&gt;There is also a family history. DunedinPoAm, the earlier version of the same idea, was validated partly in the CALERIE trial. That is not a scandal, and it does not obviously undermine the later result, but it means the clock and the trial are not strangers. How much that should matter is not settled by anything here.&lt;&#x2F;p&gt;
&lt;p&gt;“By approximately 3%” is the number everyone will remember. The same analysis gives it a translation: a 3% slower DunedinPACE “is associated with approximately 15% lower risk of death in an independent study of older adults, suggesting the CALERIE effect is clinically meaningful.” Read that literally too. Associated, not caused. Older adults, not 21-to-50-year-olds. Independent study, meaning different people. It is a bridge built from one population to another, and the bridge has not been walked across.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-did-not-produce-statistically-significant-changes-means-for-phenoage-and-grimage&quot;&gt;What “did not produce statistically significant changes” means for PhenoAge and GrimAge&lt;&#x2F;h2&gt;
&lt;p&gt;Then comes “but.” PhenoAge and GrimAge are not obscure clocks. GrimAge was trained on death, which is about as serious as training gets. In this trial, restriction did not produce statistically significant changes in either.&lt;&#x2F;p&gt;
&lt;p&gt;“Not statistically significant” does not mean “did not change.” It means whatever change happened was not distinguishable from noise with this many people. Maybe they moved a bit. Maybe they did not. The sentence declines to say, which is correct.&lt;&#x2F;p&gt;
&lt;p&gt;Here is where the post-hoc part matters. The trial was designed to study eating less; the clocks came afterward, applied to stored blood. You can imagine the conversation (this is invented): “Okay, the trial is done, the primary outcomes are published, and we have freezers full of blood. Which clocks do we run?” “All the good ones, I suppose.” That is a perfectly reasonable thing to do. It also means the result is a finding to be tested, not a test that was passed.&lt;&#x2F;p&gt;
&lt;p&gt;Now push it further, purely as an invented exercise. Suppose a future trial runs 400 clocks on its freezer blood. Statistics alone will make some of them move. Each moving clock gets its own abstract, its own press release, its own consumer kit. Clock 312 reports the control group got younger. Clock 377 says everyone is aging faster except on Tuesdays. CALERIE ran three clocks and reported all three, including the two that stayed put, which is roughly the opposite of this.&lt;&#x2F;p&gt;
&lt;p&gt;The telomere analysis of the same trial is a useful reminder of how messy one intervention can look through different instruments. By intent-to-treat, there was no significant difference in telomere attrition in the first year. In a dose analysis, more restriction went with more attrition in a methylation-based telomere estimate during the weight-loss year, and with less absolute telomere attrition during the second, weight-maintenance year. So: worse, then better, depending on the year and the ruler. It’s fine. Bodies losing weight and bodies holding weight are doing different things.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-the-calerie-result-does-and-doesn-t-prove&quot;&gt;What the CALERIE result does and doesn’t prove&lt;&#x2F;h2&gt;
&lt;p&gt;It shows that a randomized, real-world intervention moved one well-regarded aging clock by a small, measurable amount over two years. That is not nothing. Most claims in this field never get a randomized trial at all.&lt;&#x2F;p&gt;
&lt;p&gt;What it does not show is that the clock movement stands in for anything a person would feel, such as living longer or getting sick later. That is the surrogate endpoint problem, and a 2026 systematic review of geroscience intervention trials looked at it directly. It found no trial-level validation of epigenetic clocks as surrogate endpoints, no intervention-clock-outcome combination that met the criteria for meta-analysis, and certainty of evidence “rated very low across all assessed evidence bodies.” The clocks flirt with outcomes. They have not yet been shown to commit.&lt;&#x2F;p&gt;
&lt;p&gt;Anyway, the sentence says exactly what happened. Eating less for two years changed a number in some blood. That is a real result, and it is the whole result.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>How Reliable Are Consumer Biological Age Tests?</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/consumer-test-reliability/"/>
        <id>https://bioclocktest.com/consumer-test-reliability/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/consumer-test-reliability/">&lt;p&gt;Here is a small fact about consumer biological age tests: if a lab takes one tube of your DNA and runs it twice, it can hand you back two different ages. These are not two ages for two different days of your life, which would at least be interesting. They are two ages for the same tube. Elysium Health, which sells an epigenetic age test called Index, says on its science page that a single sample processed repeatedly “can result in biological age estimates that vary up to nine years.” (It says this about other people’s clocks. It reports that its own test agrees with itself within zero to 1.5 years across technical replicates.)&lt;&#x2F;p&gt;
&lt;p&gt;So the plain answer on epigenetic age test accuracy is that it depends on which clock, which lab and which kind of accuracy you mean, and there is no agreed yardstick to check any of it against. According to The Scientist, as of 2026 the field has no reference standards, so you cannot send one sample to several validated labs and get the same readout. Most kit companies also do not say which clock they use. No consumer test is approved by the U.S. Food and Drug Administration as a diagnostic; they are sold as wellness or informational products. And there are two separate kinds of wobble, and the fix for the first one does not touch the second.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-one-sample-can-come-back-with-two-ages&quot;&gt;How one sample can come back with two ages&lt;&#x2F;h2&gt;
&lt;p&gt;An epigenetic clock reads methylation, chemical tags on DNA, at particular spots called CpG sites (the chemistry has its own article on this site, so I will skip it). The Horvath pan-tissue clock uses 353 of them. The Hannum clock uses 71. Each site gets a weight, the weighted readings get added up, and out comes an age. On the job it was built for, guessing calendar age, the Horvath clock misses by about 3.6 years (on average, which hides the people it misses by more).&lt;&#x2F;p&gt;
&lt;p&gt;The trouble is that each site is measured on a chip, and every measurement carries a little error. Multiply a little error by a big weight, add up a few hundred of them, and the errors do not always cancel. Elysium reports that technical replicates of the published Horvath, Hannum and PhenoAge clocks produced deviations of three to nine years between runs. (Index is itself built on the DNAm PhenoAge biomarker, so this is a family argument.) A systematic study on a platform called TranslAGE, which tested 18 clocks, found that some showed dramatic drops in technical reliability depending on where the sample sat on the slide and which DNA extraction protocol was used.&lt;&#x2F;p&gt;
&lt;p&gt;Picture the bench. A technician loading a slide says, “Hey, this one can go in the corner, there’s room.” For most of the history of laboratory science that was a remark about where a tube fits on a piece of glass. Now it is also an age.&lt;&#x2F;p&gt;
&lt;p&gt;This is the dumb version of the problem, and I mean dumb kindly: it is a bathroom scale that reads differently depending on which tile you set it on. Everyone involved is acting sensibly. The lab packs slides efficiently because that is its job, the company wants one clean number because a report saying “somewhere between 48 and 57” is hard to sell, and the customer wants a number because a number is what they paid for.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-sophisticated-fix-principal-component-clocks&quot;&gt;The sophisticated fix: principal-component clocks&lt;&#x2F;h2&gt;
&lt;p&gt;The dumb fix would be to run every sample several times and average the results. That costs money. The sophisticated fix is to stop trusting individual sites so much (which is a lot of trust to place in a piece of glass anyway).&lt;&#x2F;p&gt;
&lt;p&gt;The general idea of a principal-component clock (a PC clock, if you want to sound like you read the paper) is that thousands of methylation sites move together in patterns. You boil those patterns down to a smaller set of summary scores and build the clock on the scores instead of the raw sites. If one site has a bad day on the chip, it is one voice in a large choir, and the choir sounds about the same. I have tried to explain principal component analysis at dinner parties, and I have mostly stopped being invited to dinner parties, so that is as far as I will go.&lt;&#x2F;p&gt;
&lt;p&gt;It works, at least on the thing it was meant to fix. In the TranslAGE study, the PC-based clocks PCGrimAge and SystemsAge remained technically reliable in all tested conditions: different slide positions, different extraction methods, same answer. The corner of the slide stopped mattering. So the problem is solved, right? You have a clock that gives the same number for the same tube.&lt;&#x2F;p&gt;
&lt;p&gt;That’s fine.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-epigenetic-age-test-accuracy-breaks-down-on-repeat-samples&quot;&gt;Why epigenetic age test accuracy breaks down on repeat samples&lt;&#x2F;h2&gt;
&lt;p&gt;Here is the catch. Nobody actually wants the same number for the same tube. People want the same number for the same person, and a person is not a tube. The TranslAGE researchers also measured biological reliability: repeated samples from the same people taken within hours of each other, before and after meals, under acute stress, across environmental exposures and over days.&lt;&#x2F;p&gt;
&lt;p&gt;Mostly the clocks did not agree with themselves. Biological reliability was markedly lower than technical reproducibility for most clocks, and most showed only moderate stability across short-interval samples. The paper puts it plainly: “clocks that were technically robust often proved biologically unreliable.” Technical reproducibility did not predict biological reliability. (Whether the PC clocks in particular held up better on the biological side is not settled by anything reported here, so I will not pretend it is.)&lt;&#x2F;p&gt;
&lt;p&gt;In some sense this is exactly where the dumb version ended up, reached by a longer and more expensive road. The dumb clock gives you a different number because the chip wobbled. The sophisticated clock gives you a different number because you ate a bagel. It is a bit like a figure skating judge who scores the same routine differently after lunch, except the routine also changed after lunch. Two draws, two ages.&lt;&#x2F;p&gt;
&lt;p&gt;What follows is invented. A customer learns all this and decides to get the best possible number. She draws blood fasting, at dawn, after nine hours of sleep, in a silent room, having cancelled every meeting for a week. She notices her number is better on Tuesdays. Soon a startup offers “draw-time optimization,” a premium tier that books your sample for the calmest twenty minutes of your month, and a rival guarantees a younger result by scheduling the draw right after a nap. Her cells are the same cells. Her calendar is different. I guess that is the sort of thing you only learn by running the experiment.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-single-numbers-get-sold-anyway&quot;&gt;Why single numbers get sold anyway&lt;&#x2F;h2&gt;
&lt;p&gt;None of this is hidden inside the field, and I feel like that is the strange part. Tony Wyss-Coray, quoted in an explainer on the science of biological age, called the consumer market “a bit like the wild West,” noting that many companies pair tests with products never tested in humans in a clinical trial. The incentives point toward one confident number, and the stuff that makes the number shaky lives in methods sections.&lt;&#x2F;p&gt;
&lt;p&gt;The AMA Journal of Ethics, in a December 2025 commentary, describes a healthy 56-year-old patient whose direct-to-consumer result said he was “biologically 65-years-old.” The commentary calls biological age a fiction, while granting that the tests have been shown to be fairly reliable indicators of risk for certain age-related diseases, and it asks companies to be clearer about how the tests work, how reliable they are and what they can and cannot show. Nine years, as it happens, is also the top of the range Elysium cites for one tube run twice. Whether that man’s nine years were signal or noise is not something anyone can check, because there is no reference standard to check it against.&lt;&#x2F;p&gt;
&lt;p&gt;The commentary also warns that results can cause psychological harm and can make age-related discrimination appear justifiable. A number with a decimal point looks like a fact. Facts get used. I would rather not think about what an HR department with a spreadsheet has planned for a column labeled “biological age.”&lt;&#x2F;p&gt;
&lt;h2 id=&quot;where-the-two-roads-meet&quot;&gt;Where the two roads meet&lt;&#x2F;h2&gt;
&lt;p&gt;Look, the engineering here is real. The industry took a tube that gave different ages depending on which corner of the slide it sat in, and built clocks that give that tube the same age every time. Then it asked the same customer for a second tube, taken after breakfast, and got a different age.&lt;&#x2F;p&gt;
&lt;p&gt;The tube was never the hard part.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>DNA Methylation: The Chemical Tags Clocks Read</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/dna-methylation/"/>
        <id>https://bioclocktest.com/dna-methylation/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/dna-methylation/">&lt;p&gt;Your DNA has small things stuck to it, and an industry now reads them to tell you how old you are. The stuck-on things are methyl groups, and the process is DNA methylation, which Wikipedia describes as “a biological process by which methyl groups are added to the DNA molecule, which can change the activity of a DNA segment without changing the sequence.” The letters stay the same. Some of them just wear hats.&lt;&#x2F;p&gt;
&lt;p&gt;The hats go in particular places. In mammals, methylation is found almost exclusively at CpG dinucleotides (a C followed by a G along the strand), with the cytosines on both strands usually methylated. Which sites wear hats differs between young people and old people, and that difference is the raw material of every epigenetic clock: first-generation clocks such as the Horvath clock were trained, as The Scientist puts it, “by comparing DNA methylation signatures of young and old individuals to predict chronological age.”&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-dna-methylation-is-in-plain-chemistry&quot;&gt;What DNA methylation is, in plain chemistry&lt;&#x2F;h2&gt;
&lt;p&gt;A cell attaches a small chemical tag, a methyl group, to one of the letters of its DNA. The tagged stretch can then act differently, while the spelling never changes. As of 2016, two bases were known to take natural, enzymatic methylation: adenine and cytosine. The clocks live on the cytosine side, at those CpG sites.&lt;&#x2F;p&gt;
&lt;p&gt;This is a little weird, if you think about it. Your genome is the most famous unchangeable document you own, and the part the biological-age business wants to read is the annotation in the margins. (Like buying a used textbook for the highlighting.)&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-one-tag-becomes-an-age-one-reasonable-step-at-a-time&quot;&gt;How one tag becomes an age, one reasonable step at a time&lt;&#x2F;h2&gt;
&lt;p&gt;Here is a progression I made up, to show how you get from a hat to a number. It is invented, though you will see it does not stay invented for long.&lt;&#x2F;p&gt;
&lt;p&gt;Step one: look at a single CpG site in a single cell. It is tagged or it isn’t. That is a fact about one molecule and nothing else, and that’s fine.&lt;&#x2F;p&gt;
&lt;p&gt;Step two: look at that site across all the cells in a tube of blood and report the share that are tagged. Still reasonable. A blood sample is a crowd, and crowds have averages.&lt;&#x2F;p&gt;
&lt;p&gt;Step three: notice that at some sites, the share tends to differ between old donors and young donors. Also reasonable. Lots of things differ between old and young people (knees, ringtones, opinions about fonts, etc.), and some of those differences turn out to be useful.&lt;&#x2F;p&gt;
&lt;p&gt;Step four: measure hundreds of thousands of sites at once, because the equipment lets you. Step five: hand all of it to a machine-learning algorithm and ask which weighted combination best predicts the donor’s age. Step six: print the answer in an app, to one decimal place. (The decimal place is doing a lot of work.)&lt;&#x2F;p&gt;
&lt;p&gt;Step seven, to push it until it breaks: skip the blood. Put a piece of tape on someone’s face, peel it off, and read their age from whatever comes off with it. That would be the absurd endpoint, I suppose.&lt;&#x2F;p&gt;
&lt;p&gt;It is not hypothetical. A study of epigenetic age predictors for human skin reports that “Skin biological age can be measured from a person’s face using non-invasive tape-stripping and DNA methylation clocks.” Steps one through six are also just how the field works; clocks do use machine learning on genome-wide methylation data.&lt;&#x2F;p&gt;
&lt;p&gt;Said as plainly as possible, the method is: count hats at a lot of spots, multiply each count by a weight, add it all up, and call the total an age. The official name for this is an epigenetic clock.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-arrays-that-read-450-000-sites-at-once&quot;&gt;The arrays that read 450,000 sites at once&lt;&#x2F;h2&gt;
&lt;p&gt;The equipment in step four is usually a methylation array, a chip that reports methylation levels at a fixed menu of CpG sites. The Hannum clock, published in 2013, came from whole-blood methylation profiling of 482 people, at more than 450,000 CpG markers. Another study, working from roughly 470,000 CpGs in whole blood, built clocks that predict age with R² scores above 0.73.&lt;&#x2F;p&gt;
&lt;p&gt;The chips have names like printer cartridges. DunedinPACE can be computed from Illumina 450k, EPIC1 or EPIC2 array data, and a study of childhood lead exposure read ten different clocks off Infinium MethylationEPIC BeadChips. (I explain things for a living and still had to read “BeadChip” three times.)&lt;&#x2F;p&gt;
&lt;p&gt;Arrays are not the only reader. A forensic study tried Oxford Nanopore’s PromethION 2 sequencer on samples under 100 ng of DNA, and found that at low read depth, sites came back with values of exactly zero or exactly one. This makes sense. If you only see one copy, it has a hat or it doesn’t, and you are back at step one.&lt;&#x2F;p&gt;
&lt;p&gt;A separate coverage-aware evaluation of Nanopore methylation callers found that at single-read coverage, estimates were “unstable and strongly method-dependent” across all four callers it tested. Somewhere a technician says, “Hey, this one’s a one,” and somewhere downstream a model treats that as a site that is 100 percent tagged across a whole crowd of cells. (The forensic team did find that a simple linear correction significantly improved accuracy, which is a nice result and also an admission.)&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-methylation-changes-with-age-and-what-is-not-settled&quot;&gt;How methylation changes with age, and what is not settled&lt;&#x2F;h2&gt;
&lt;p&gt;The sources here establish two things. Methylation patterns shift with age enough to predict age. And the relationship is not constant: researchers behind an algorithm called iTARGET describe “Epigenetic Correlation Drift,” the way the link between methylation and age changes across life stages. Their answer is to sort profiles into age groups first and then fit a separate model for each group.&lt;&#x2F;p&gt;
&lt;p&gt;Which particular sites gain or lose hats, why, and whether those changes cause aging or merely accompany it: the facts gathered for this article do not settle that. I will not pretend they do.&lt;&#x2F;p&gt;
&lt;p&gt;There is also the crowd problem. Blood is a mix of cell types, and different cell types carry different hat patterns, so a shift in the mix can look like a shift in age. First-generation clocks put people living with HIV at roughly 5 to 7 years biologically older than HIV-negative people of the same age, though that gap may be confounded by HIV-related changes in immune cell composition. Is that aging, or a different crowd? It is not settled.&lt;&#x2F;p&gt;
&lt;p&gt;Saliva has the same problem in a bigger way: a clock trained on blood, applied to saliva without correction, can produce substantial errors, because the cell types in saliva carry fundamentally different methylation signatures.&lt;&#x2F;p&gt;
&lt;p&gt;The builders know all this. DunedinPACE, per its paper in eLife, was trained on a methylation dataset “restricted to exclude probes with low test-retest reliability.” That is, they cut the sites that told one story the first time and a different story when asked again the next week. Many people run their social lives on the same policy.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-everyone-keeps-adding-sites-and-decimal-places&quot;&gt;Why everyone keeps adding sites and decimal places&lt;&#x2F;h2&gt;
&lt;p&gt;Each party here is behaving sensibly. A lab whose chip returns 450,000 numbers per run would be strange to use only twelve of them. A company cannot hand a customer 450,000 numbers, because the customer asked how old they are, and the honest answer to that question is a single figure. And the customer wants a single figure because a single figure is what the question was.&lt;&#x2F;p&gt;
&lt;p&gt;The researchers have been reasonably upbeat. Podolskiy and Gladyshev, in a review submitted on 14 November 2018, wrote that methylation clocks “provide unmatched accuracy in assessing the biological age of both humans and animal models of aging.” The same review dealt with bounds on clocks’ out-of-sample performance, which is the part where you admit a model has limits.&lt;&#x2F;p&gt;
&lt;p&gt;On the other hand, a 2025 meta-analysis of 24 studies covering 28,325 participants found that raw methylation age, unadjusted for chronological age, showed no association with frailty. I guess that is what you would expect from a number trained mainly to reproduce your birthday. Anyway, the other articles on this site cover what the clocks predict; this one is about the hats.&lt;&#x2F;p&gt;
&lt;p&gt;The letters, again, still haven’t changed. The hats now come off on a strip of tape and go home with a decimal point.&lt;&#x2F;p&gt;
&lt;p&gt;Your sequence, the part you were born with and will die with, has become the boring part.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>Do Ageing Clocks Predict Disease and Death?</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/do-aging-clocks-predict-death/"/>
        <id>https://bioclocktest.com/do-aging-clocks-predict-death/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/do-aging-clocks-predict-death/">&lt;p&gt;Researchers have spent years taking blood from people, reading chemical tags off their DNA, and then waiting to see who dies. It mostly works, in the narrow way it was designed to. In blood-based studies, the epigenetic clock mortality evidence shows that clocks like GrimAge are associated with higher risk of death and disease, sometimes even after the researchers account for other things they already knew about the participants. The effects are real. In the most carefully adjusted study here, they are also small.&lt;&#x2F;p&gt;
&lt;p&gt;What the evidence does not show is that a clock can substitute for actually waiting, that moving your clock moves your odds, or that a cheek swab means what a blood draw means. A 2026 systematic review of intervention trials found no validated case of the first. The easiest way to see why is to build the idea up one reasonable step at a time, until it stops being reasonable.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-a-clock-that-only-guesses-your-age-is-good-for&quot;&gt;What a clock that only guesses your age is good for&lt;&#x2F;h2&gt;
&lt;p&gt;Step one is modest. You take DNA methylation (the chemistry has its own article on this site) and fit a model that guesses how old someone is. The Horvath pan-tissue clock does this with 353 CpG loci and lands within about 3.6 years on average across many tissue types. The Hannum clock, built from the whole blood of 482 people, uses 71.&lt;&#x2F;p&gt;
&lt;p&gt;This is an elaborate machine for producing a number already printed on your driver’s license.&lt;&#x2F;p&gt;
&lt;p&gt;Nobody wants it for that, presumably. The interesting part is the miss. If the clock says 58 and the license says 50, maybe the extra eight years mean something, and maybe the body is running ahead of the calendar. That is the premise of the whole field, and it is a sensible premise.&lt;&#x2F;p&gt;
&lt;p&gt;The trouble is that a clock graded only on hitting the calendar treats those eight years as an error it would rather not have made. CD Genomics describes first-generation clocks as “better at capturing normal aging than pathological aging related to disease and mortality risk.” They were built to be right about your birthday. Being right about your death was never on the rubric.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-happens-when-you-train-the-clock-on-death&quot;&gt;What happens when you train the clock on death&lt;&#x2F;h2&gt;
&lt;p&gt;Step two: change the rubric. Second-generation clocks, PhenoAge and GrimAge, swapped the training target from chronological age to clinical phenotypes, disease incidence and all-cause mortality (GrimAge is the one literally trained on death, and it gets its own article here).&lt;&#x2F;p&gt;
&lt;p&gt;So now the clock predicts death. Well, no. It predicts death in the data it was fitted on, which is in some sense just the exam it already saw. GrimAge’s relationship with its training cohort is, professionally speaking, blurry: they were introduced early, spent a lot of late nights alone with the same spreadsheet, and now each one finishes the other’s sentences. The stranger and more precise version is that a mortality clock, on the day it is born, is a very elaborate summary of deaths that have already happened.&lt;&#x2F;p&gt;
&lt;p&gt;The only question worth asking is whether it works on deaths that haven’t happened yet. That is prospective validation, and it mostly consists of patience. You measure living people and write down what the clock says. Then you wait.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-the-epigenetic-clock-mortality-numbers-say&quot;&gt;What the epigenetic clock mortality numbers say&lt;&#x2F;h2&gt;
&lt;p&gt;Step three is the waiting. The PhenoAge validation in NHANES IV had mortality follow-up spanning 12.6 years, using National Death Index records through December 31, 2011. (I find it hard to wait for a large PDF to finish loading, which is part of why I write about cohort studies instead of running them.)&lt;&#x2F;p&gt;
&lt;p&gt;The result usually comes out as a hazard ratio. A hazard ratio of 1.40 means that at any moment during follow-up, one group is dying at 1.4 times the rate of the comparison group. When it is stated “per standard deviation,” it means each step of ordinary spread in the clock reading multiplies the rate by that amount.&lt;&#x2F;p&gt;
&lt;p&gt;In the Head and Neck 5000 cohort, made up of people already diagnosed with oropharyngeal cancer, accelerated GrimAge came with a hazard ratio for all-cause mortality of 1.40 per standard deviation, with a 95% confidence interval of 1.06 to 1.83. Accelerated Hannum age came in at 1.30 (1.07 to 1.57). The confidence interval is the range the study can’t rule out, and for GrimAge it runs from a 6% bump to an 83% bump. That is a range you could park a bus in.&lt;&#x2F;p&gt;
&lt;p&gt;It isn’t only methylation. Deep-learning clocks built from routine blood markers in NHANES produced mortality hazard ratios of 1.059 and 1.039 for their accelerated-aging measures, and were linked to deaths from cardiovascular and cerebrovascular disease and cancer. A proteomic clock built in the UK Biobank from 2,923 proteins in 53,021 people showed stronger age-adjusted associations with mortality and 18 diseases than PhenoAge, BioAge or telomere length. Everybody’s clock predicts death a little. The competition is over the margins.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;whether-the-clock-beats-the-usual-risk-factors&quot;&gt;Whether the clock beats the usual risk factors&lt;&#x2F;h2&gt;
&lt;p&gt;Step four is the one medicine actually cares about. Doctors already have standard risk factors (the stuff a checkup already measures), and a new test is useful only if it adds something those don’t. Statisticians check this by adjusting: they put the known factors into the model alongside the clock and see what is left.&lt;&#x2F;p&gt;
&lt;p&gt;In the MESA study’s fully adjusted models, GrimAge was still independently associated with higher risk of composite cardiovascular disease, with a hazard ratio of 1.05 (1.02 to 1.08). The interval stays above 1, so it isn’t nothing. It is 5%.&lt;&#x2F;p&gt;
&lt;p&gt;Five percent.&lt;&#x2F;p&gt;
&lt;p&gt;Look, there are two ways to read that, and the facts here don’t pick one. In the first, after you remove everything clinicians already measure, there is still something in the methylation that knows about the heart, which is strange and interesting. In the second, the clock mostly re-measures what the doctor already had, and 1.05 is the crumbs left over once the doctor’s numbers have eaten. Both readings fit the same number.&lt;&#x2F;p&gt;
&lt;p&gt;Other studies point in the same direction without settling it. In men with hypertension, ischemic heart disease, obesity or dyslipidemia, only GrimAge and DunedinPACE picked up the disease signal; none of 19 first-generation clocks did. A commentary in the AMA Journal of Ethics calls epigenetic clock tests “fairly reliable indicators” of risk for certain age-related diseases, in the same piece that calls biological age a fiction. I feel like both of those can be true.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-a-clock-can-t-yet-stand-in-for-the-outcome&quot;&gt;Why a clock can’t yet stand in for the outcome&lt;&#x2F;h2&gt;
&lt;p&gt;Step five is where the road gets steep. Following people for 12.6 years is slow and expensive. It would be much nicer to run a short trial, see whether a treatment lowers the clock, and call that a win. Medicine uses surrogate endpoints like this all the time, once the surrogate has earned it.&lt;&#x2F;p&gt;
&lt;p&gt;Here is an invented version, pushed to the end. A company runs a trial, and its pill lowers DunedinPACE (which the Moffitt and Caspi lab describes as the clock most sensitive to intervention). A sponsor tells the investor call, “Hey guys, the cohort got 0.3 years younger this quarter.” A rival pill lowers it more. An insurer prices premiums off the clock. A dating app lists it next to height. At no point does anyone check whether fewer people died, because checking would take 12.6 years, and the whole point was not to.&lt;&#x2F;p&gt;
&lt;p&gt;The real world has not gotten there, and the reason is boring. A 2026 systematic review of 26 geroscience intervention studies found that changes in clocks and changes in clinical outcomes lined up inconsistently, and rated the certainty of evidence “very low” across every body of evidence it assessed. It identified no trial-level surrogate validation, and concluded that clocks should be treated as exploratory or adjunctive biomarkers. Moving the clock has not been shown to move the thing it predicts.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-far-down-this-road-the-real-world-has-gotten&quot;&gt;How far down this road the real world has gotten&lt;&#x2F;h2&gt;
&lt;p&gt;Steps one through four are done, in blood, at the level of large groups. The Scientist notes that the link to disease and mortality “has not been established for saliva-based clocks,” and that a blood-trained clock applied to saliva without correction can produce substantial errors. As of 2026 the field has no reference standards, so one sample sent to different validated labs need not come back with the same age.&lt;&#x2F;p&gt;
&lt;p&gt;Then there is stability. Nearly all of 18 clocks tested on the TranslAGE platform reproduced well on replicate assays, but reliability across samples taken hours apart, before and after meals or under stress was markedly lower for most, and technical robustness did not predict biological reliability. Elysium Health, on a page explaining the science behind its Index product, reports that technical replicates of Horvath, Hannum and PhenoAge deviated by three to nine years. It is a bathroom scale that agrees with itself twice in a row and then disagrees after lunch.&lt;&#x2F;p&gt;
&lt;p&gt;None of this is dumb. A hazard ratio of 1.05 survives that kind of noise because it is averaged over thousands of people, and I assume that is exactly why the cohort studies work. A single reading for a single person gets averaged over nobody.&lt;&#x2F;p&gt;
&lt;p&gt;So the clocks do predict death, for crowds. For one person, the only test fully validated against death is still death.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>DunedinPACE: Measuring the Speed of Ageing</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/dunedinpace/"/>
        <id>https://bioclocktest.com/dunedinpace/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/dunedinpace/">&lt;p&gt;There is a blood test that does not tell you how old you are. It tells you how fast you are getting older, which is a different number and a more personal one.&lt;&#x2F;p&gt;
&lt;p&gt;The test is DunedinPACE, a DNA methylation measure built from one group of people born in 1972 and 1973, whose bodies were measured over and over for two decades. Most epigenetic clocks read chemical tags on your DNA and output an age in years. This one reads the same kind of tags and outputs a rate: how quickly the body’s systems are wearing down, relative to the calendar. Researchers sometimes call it a “third generation” clock because, unlike the generations before it, it was trained on change over time rather than on a snapshot.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-the-clock-measures-instead-of-an-age&quot;&gt;What the clock measures instead of an age&lt;&#x2F;h2&gt;
&lt;p&gt;The older clocks (the site covers them elsewhere) were mostly built by taking blood from people of many ages and teaching an algorithm to guess each person’s age from their methylation. That works. It also mostly teaches the algorithm to be a very expensive birth certificate.&lt;&#x2F;p&gt;
&lt;p&gt;The pace clock was built differently. The Scientist describes third-generation clocks as “trained on differences in multi-organ system deterioration between individuals of the same age.” The target was never “how old is this person.” It was “among people who are exactly the same age, who is falling apart faster.” Papers on Alzheimer’s markers put it plainly: the measure captures the pace of biological aging rather than a static biological age estimate.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-objection-isn-t-a-speed-just-two-ages&quot;&gt;The objection: isn’t a speed just two ages?&lt;&#x2F;h2&gt;
&lt;p&gt;Here is the sensible reaction. A speed is distance divided by time. If an age clock says your body is 40 now and 42 a year from now, you aged two years in one. Why do you need a separate clock for that? Just take two readings and subtract.&lt;&#x2F;p&gt;
&lt;p&gt;To be fair, that is how speed works. But consider an invented case. A man (call him Gary, he is not real) takes an age test every month. Any measurement wobbles a little, so his readings go 51, 47, 53, 49. Subtracting, he aged four years in January, reversed four years in February, aged six in March. By summer he is Benjamin Button with a mortgage (Gary would like a word with the lab). Subtraction amplifies wobble.&lt;&#x2F;p&gt;
&lt;p&gt;The pace clock does something stranger than subtraction. It produces a speed from a single blood draw. It is a radar gun that works on a photograph. That sounds impossible, and in some sense the whole design of the study exists to make it possible.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-the-dunedin-study-built-a-speedometer&quot;&gt;How the Dunedin Study built a speedometer&lt;&#x2F;h2&gt;
&lt;p&gt;The trick is the cohort. Everyone in the Dunedin Study was born in the same 1972 to 1973 window, so at every visit they were all the same age. That removes the usual problem. If two 38-year-olds differ, age cannot be the reason, because they have the same one.&lt;&#x2F;p&gt;
&lt;p&gt;So the researchers measured. The eLife paper describes tracking “within-individual decline in 19 indicators of organ-system integrity across four time points spanning two decades,” at ages 26, 32, 38 and 45. Each person got a personal slope: how fast their 19 indicators slid over those years. That slope is the Pace of Aging. Then the algorithm was trained to find the methylation pattern in a single blood sample that matched each person’s slope.&lt;&#x2F;p&gt;
&lt;p&gt;I assume the logistics were less elegant than the math. “Great, thanks for coming in, see you when you’re 32” is a sentence somebody presumably said to a 26-year-old, and then meant it. (Richie Poulton, whose role in the Dunedin Study is reflected in his co-authorship of the paper, was one of the people meaning it.)&lt;&#x2F;p&gt;
&lt;p&gt;This was the second attempt. A 2020 paper by Belsky and colleagues first showed that pace could be estimated from a methylation blood test, using a version called DunedinPoAm, which modeled change in 18 biomarkers over ages 26 to 38. The later version added a 19th indicator and a fourth time point. Same idea, longer road.&lt;&#x2F;p&gt;
&lt;p&gt;How well does the photograph match the road? The pace score correlates with the 20-year Pace of Aging at r = 0.78. The paper also checked it against functioning at age 45: physical change between the age-38 and age-45 visits, and cognitive change all the way from an age-13 baseline. Thirty-two years. That is a long time to wait for a data point. The clock was then tested in five additional datasets, and its developers’ lab page says it has been validated in more than 65 large cohorts and reported in more than 300 publications, which is more papers than I have read about anything, including the things I write about.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-a-faster-or-slower-dunedinpace-has-been-linked-to&quot;&gt;What a faster or slower DunedinPACE has been linked to&lt;&#x2F;h2&gt;
&lt;p&gt;The Moffitt &amp;amp; Caspi lab page calls it “the most sensitive epigenetic clock to intervention.” The main evidence for responsiveness is the CALERIE trial, where 25% caloric restriction over 2 years slowed the pace by about 3%, in a post-hoc analysis, while PhenoAge and GrimAge did not change significantly. That is a modest number, which is part of why it is interesting. Separately, a 3% slower pace has been associated with about 15% lower risk of death in an independent study of older adults. Whether the first number buys you the second is not settled. (Those are two different groups of people, which matters.)&lt;&#x2F;p&gt;
&lt;p&gt;Then there is the number that stops you. In a study of people with well-controlled HIV, all 17 individuals who died had a baseline value greater than 1. All 17, every one of them above the same line, measured before anything happened. It is a small group, and a small group can line up neatly by chance. But it is the kind of tidy that makes a statistician put down their coffee.&lt;&#x2F;p&gt;
&lt;p&gt;Other results are quieter. In the WHIMS study, higher baseline pace went with faster rises over 15 years in plasma markers tied to Alzheimer’s disease. In men with heart and metabolic problems, GrimAge and the pace clock picked up those diseases while none of 19 first-generation clocks did. In Add Health, lifestyle and sociodemographic factors showed up more often in clocks like this one than in clocks trained on age alone. In a Parkinson’s cohort that computed seven different clocks, the pace clock was the one linked to rigidity in women. (Running seven clocks on the same blood is the research version of asking every friend for directions.)&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-to-read-a-pace-number-and-what-is-not-settled&quot;&gt;How to read a pace number, and what is not settled&lt;&#x2F;h2&gt;
&lt;p&gt;Look, a speed is a slippery thing to hold. A frailty meta-analysis found that a higher pace went with more frailty at a single point in time (5 studies, 7,895 people, a standardized effect of 0.10), but with heterogeneity of 91%, which is the statistical way of saying the studies sort of disagreed with each other. Followed over time, the pace did not significantly predict rising frailty. GrimAge did, barely.&lt;&#x2F;p&gt;
&lt;p&gt;So what is one reading worth? I don’t know, and the literature does not quite know either. It is an estimate, made from one moment, of a slope that took the original cohort twenty years to draw. That is impressive. It is also an inference about the past being sold as a reading of the present.&lt;&#x2F;p&gt;
&lt;p&gt;The clock has also left the lab. TruDiagnostic’s TruAge combines it with two other clocks, as part of a report covering 64 biomarkers. A clock goes from a citation, to a collaboration, to a product page, and the line between a measuring instrument and a dashboard you check over breakfast gets a little blurry.&lt;&#x2F;p&gt;
&lt;p&gt;Other fields want the same trick, for instance Brain-PACE, an MRI method posted in September 2026 that estimates the pace of brain aging from paired scans. Everyone would like a speedometer.&lt;&#x2F;p&gt;
&lt;p&gt;Most clocks read the odometer. This one reads the speedometer, off one blood sample, using twenty years of other people’s roads. It does not know where you are going.&lt;&#x2F;p&gt;
&lt;p&gt;It knows how fast.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>How an Epigenetic Clock Is Built</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/epigenetic-clock/"/>
        <id>https://bioclocktest.com/epigenetic-clock/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/epigenetic-clock/">&lt;p&gt;Here is how you build a machine that tells people how old they are. You collect blood from a few hundred people whose ages you already know. You measure a chemical mark at hundreds of thousands of spots on their DNA. Then you ask a computer to find the combination of spots that best guesses the ages you already knew, and you publish the guesses.&lt;&#x2F;p&gt;
&lt;p&gt;That, more or less, is an epigenetic clock. The chemical mark is DNA methylation (other articles on this site handle the chemistry), the spots are called CpG sites, and the guessing is usually done with a statistical method called elastic net regression. UCSF’s summary says the sites in the two original clocks, Horvath and Hannum, “were selected using elastic net regression aimed at producing the most accurate estimate of chronological age.” The Hannum clock was built from the whole blood of 482 people and uses 71 CpG sites. The Horvath clock uses 353, and according to CD Genomics it guesses chronological age across many tissue types with a mean absolute error of about 3.6 years.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-train-a-clock-to-guess-something-you-already-know&quot;&gt;Why train a clock to guess something you already know&lt;&#x2F;h2&gt;
&lt;p&gt;A sensible person hears this and objects. The number the clock was trained to predict is printed on your driver’s license. If the clock were perfect, it would tell you your birthday, which you presumably knew. Building a very expensive instrument to reproduce a fact available on any government ID (and on most birthday cakes) seems dumb.&lt;&#x2F;p&gt;
&lt;p&gt;The objection is correct. A perfect first-generation clock would be useless. The clock is worth something only to the extent that it misses, and misses in a patterned way. To make up a case: if it says 52 for someone who is 47, the hope is that the extra five years say something about that person’s body rather than something about the lab. So a first-generation clock is faithful to chronological age, mostly, and the interesting stuff lives in its little lapses of fidelity.&lt;&#x2F;p&gt;
&lt;p&gt;This is a strange way to build a measuring instrument. Normally you build a thermometer to measure temperature, and its errors are errors. Here the errors are the product, and the training procedure spends all its effort trying to eliminate the product. It just can’t quite manage it.&lt;&#x2F;p&gt;
&lt;p&gt;Think of the guy at the state fair who guesses your age for a stuffed animal. If he were perfect, nobody would learn anything except that he is good at his job. But if he reliably guesses high for people who have had a rough decade, his mistakes become sort of interesting. The clock is that guy, working from hundreds of clues instead of your hairline. We only care when he’s wrong.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-elastic-net-picks-a-few-hundred-sites-out-of-hundreds-of-thousands&quot;&gt;How elastic net picks a few hundred sites out of hundreds of thousands&lt;&#x2F;h2&gt;
&lt;p&gt;The mechanical problem is that you have far more variables than people. CD Genomics describes first-generation clocks screening for the CpG loci with the strongest linear relationship to age “from among hundreds of thousands of loci.” Hannum had 482 people. If you hand an ordinary regression hundreds of thousands of inputs and a few hundred answers, it will fit the training data perfectly, by memorizing it, and then be useless on anyone new.&lt;&#x2F;p&gt;
&lt;p&gt;Penalized regression fixes this by charging the model for every weight it uses. Elastic net blends two kinds of charge: one that pushes weak weights all the way to zero, and one that shrinks weights and lets groups of correlated sites share the credit. The model keeps a site only if that site earns more in accuracy than it costs in penalty. Almost everything gets a weight of zero. (This is also roughly how I write these articles, except the regression is more disciplined about what it cuts.) A paper on feature selection notes that the most common approach runs elastic net on all measured CpG sites “without first identifying specific features or CpGs of interest.”&lt;&#x2F;p&gt;
&lt;p&gt;So the 353 sites in the Horvath clock are the 353 places on the genome that matter for aging. Actually, no. They are the set that, given that training data and that penalty, did the cheapest job of guessing age. The regression has no opinion about biology (it is very sure about arithmetic). Swap in a different sample, and I assume some sites drop out and their correlated neighbors step in to do the same job.&lt;&#x2F;p&gt;
&lt;p&gt;Picture the lab meeting. “Okay, it zeroed out basically the whole genome and kept these. Does anyone know what they do?” “Not really. But the error is 3.6 years.” That is a perfectly reasonable exchange, in some sense, because the method never promised an explanation. It promised a fit.&lt;&#x2F;p&gt;
&lt;p&gt;The number of sites is not sacred either. The same feature-selection study reports that its most predictive clock used only 35 CpG sites and reached an R² of 0.87, and that a clock built from just the five sites that showed up most often across its clocks reached 0.83. A skin clock called MitraSolo is built on single CpG sites, while a sibling called MitraCluster uses CpG regions. Hundreds of sites, dozens, a handful. They all tell time.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-the-second-and-third-generations-changed-the-target&quot;&gt;How the second and third generations changed the target&lt;&#x2F;h2&gt;
&lt;p&gt;Once you notice that the errors are the product, the obvious move is to stop training on chronological age and train on the thing you actually wanted. That is what the later generations did. CD Genomics says second-generation clocks like PhenoAge and GrimAge changed the training objective “from predicting chronological age to predicting clinical phenotypes, incidence of disease, and all-cause mortality associated with aging.” The machinery stayed about the same. The answer key changed.&lt;&#x2F;p&gt;
&lt;p&gt;The third generation changed the answer key again. The Scientist describes DunedinPACE as trained on differences in multi-organ system deterioration between individuals of the same age. The Moffitt and Caspi lab, which built it, says the “third generation” label exists because earlier clocks were not derived from longitudinal data on physiological decline in same-age people. Its output is a rate of aging rather than an age, and the lab describes it as the clock most responsive to interventions across studies.&lt;&#x2F;p&gt;
&lt;p&gt;Look, “generation” here sounds like it means better chips, the way phones get generations. It mostly means a different target variable. First generation: guess the birthday. Second: guess the health outcomes and the death. Third: guess how fast someone is declining compared with people born the same year. Each generation is a more direct admission that the birthday was never the point. (Individual clocks get their own articles here, so I will leave them there.)&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-the-clocks-end-up-measuring-along-the-way&quot;&gt;What the clocks end up measuring along the way&lt;&#x2F;h2&gt;
&lt;p&gt;Because the regression keeps whatever predicts, it can keep things nobody meant to measure. A preprint called GIA mapped the CpG sites from 13 published clocks to places where methylation is known to be regulated by inherited genetic variants in blood. Clock sites were enriched for those, 62.7% versus 39.6% for comparable unused sites on the same array, an odds ratio of 2.57.&lt;&#x2F;p&gt;
&lt;p&gt;One reading is that clocks are partly measuring the genes you were born with, which would be an odd ingredient in a gauge of how you have aged since. The other reading is that genetically regulated sites are simply steadier signals, and a penalized regression hunting for reliable predictors would naturally favor steady ones. Those lead to different conclusions about what a clock reading means. The preprint’s numbers do not choose between them.&lt;&#x2F;p&gt;
&lt;p&gt;Other passengers show up too. Research on people living with HIV finds first-generation clocks put them roughly 5 to 7 years older than HIV-negative peers, while noting that this may be confounded by changes in immune cell composition. The Scientist reports that a blood-trained clock applied to saliva without correction can produce substantial errors. Elysium Health (which sells its own test) says technical replicates of the Horvath, Hannum and PhenoAge clocks produced deviations of three to nine years. A study of 18 clocks on the TranslAGE platform found that clocks which were technically robust often proved biologically unreliable across samples taken hours apart.&lt;&#x2F;p&gt;
&lt;p&gt;And there is no shared ruler. The Scientist says the field has not achieved reference standards, so “it is not possible to send a sample to different validated labs and obtain the same biological age readout.” A 2026 systematic review concluded clocks should be treated as exploratory or adjunctive biomarkers, not validated surrogate endpoints. The AMA Journal of Ethics went further and called “biological age” a fiction. I feel like the builders would take that less hard than it sounds, since they started with a number they already knew.&lt;&#x2F;p&gt;
&lt;p&gt;The clock was built to be right about your age. The useful part is where it’s wrong.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>GrimAge: The Clock Trained on Death</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/grimage/"/>
        <id>https://bioclocktest.com/grimage/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/grimage/">&lt;p&gt;I spent an unreasonable part of this week reading one sentence from a lab-services company’s explainer page the way a lawyer reads an indemnity clause, because it turns out to be the cleanest description of GrimAge available. If you typed grimage into a search box, here is the plain answer: GrimAge is an epigenetic clock, meaning a formula that reads chemical tags on DNA (usually from a blood sample) and outputs a number in years. It was published by Lu and colleagues in 2019, with Steve Horvath among its developers. What makes it different is what it was trained to predict. Not birthdays.&lt;&#x2F;p&gt;
&lt;p&gt;Death.&lt;&#x2F;p&gt;
&lt;p&gt;A GrimAge higher than someone’s calendar age means their blood methylation looks like the blood of people who died sooner. In a nationally representative U.S. cohort (NHANES) followed for 17.5 years, it was the strongest mortality predictor among the clocks compared. What it has not been shown to be, according to a 2026 systematic review, is a validated target that an intervention can push down with a known payoff in years of life. Those are different claims, and most of what follows is about the distance between them.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-sentence-and-what-a-training-objective-is&quot;&gt;The sentence, and what a training objective is&lt;&#x2F;h2&gt;
&lt;p&gt;Here is the sentence, from CD Genomics: “Second-generation clocks such as PhenoAge and GrimAge shifted the training objective from predicting chronological age to predicting clinical phenotype, disease incidence, and all-cause mortality.” It reads like a software changelog. It is actually the whole story, so let’s take it clause by clause.&lt;&#x2F;p&gt;
&lt;p&gt;Start with “training objective.” The plain version: you collect a big pile of blood samples, you attach an answer to each one, and you let a computer hunt for the combination of methylation sites that best guesses the answer. Whatever you write in the answer column, the computer learns to guess. That column is the training objective. (It is the least glamorous phrase in the sentence and the only one doing any work.)&lt;&#x2F;p&gt;
&lt;p&gt;The point is that a clock has no opinion about aging. It has an answer key. Change the key and you get a different clock out of the same DNA.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;from-chronological-age-which-was-the-easy-part&quot;&gt;From chronological age, which was the easy part&lt;&#x2F;h2&gt;
&lt;p&gt;“From predicting chronological age.” This is the first generation. The Scientist describes those clocks, Horvath’s included, as trained “by comparing DNA methylation signatures of young and old individuals to predict chronological age.” The Horvath pan-tissue clock uses 353 CpG sites and lands within about 3.6 years of a person’s real age on average, across many tissue types. As engineering, that is impressive.&lt;&#x2F;p&gt;
&lt;p&gt;As a health test, it has an obvious problem. If the answer key is your age, the perfect clock tells you your age. You already knew that. It’s a very expensive calendar.&lt;&#x2F;p&gt;
&lt;p&gt;People noticed. The same CD Genomics page says the purely chronological objective makes first-generation clocks better at capturing normal aging than the pathological kind tied to disease and death. The data agree in their own way. In the Add Health study of younger U.S. adults, sociodemographic and lifestyle factors showed up more often in clocks trained on health outcomes (PhenoAge, GrimAge, DunedinPACE) than in Horvath. And in a study of men with hypertension, heart disease, obesity and dyslipidemia, none of 19 first-generation clocks registered those conditions. GrimAge and DunedinPACE did.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-grimage-is-trained-to-predict-read-literally&quot;&gt;What GrimAge is trained to predict, read literally&lt;&#x2F;h2&gt;
&lt;p&gt;“To predicting clinical phenotype, disease incidence, and all-cause mortality.” The last item is the one that earned the name. “All-cause” means what it says: the clock does not care what killed anyone (heart attack, cancer, falling piano), only that they died. This is basically the premise of the Final Destination films, minus the set pieces and plus a great deal of regression.&lt;&#x2F;p&gt;
&lt;p&gt;So GrimAge predicts when you will die. That’s not quite right, and the precise version is stranger. It does not produce a date. It scores how much someone’s blood resembles the blood of people who died sooner, and the score comes out in years (presumably because “62” fits on a results page better than a hazard ratio does). A GrimAge “age” is a mortality score, expressed in years.&lt;&#x2F;p&gt;
&lt;p&gt;As a mortality score, it is pretty good. In the NHANES cohort, each five years of GrimAge acceleration carried a hazard ratio of 1.50 for death from any cause, against 1.15 for PhenoAge, 1.17 for Hannum and 1.12 for Horvath. GrimAge acceleration also specifically predicted cardiovascular death (hazard ratio 1.46). In people with well-controlled HIV, positive acceleration on GrimAge versions one and two was each linked to more than double the risk of serious non-AIDS events. Train a clock on death and it gets good at death. Lol, of course it does.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-smoking-stand-in-and-grimage2&quot;&gt;The smoking stand-in and GrimAge2&lt;&#x2F;h2&gt;
&lt;p&gt;Two pieces of GrimAge lore come up constantly, and on both the honest answer is shorter than the lore.&lt;&#x2F;p&gt;
&lt;p&gt;The first is smoking. GrimAge has a reputation for containing a methylation-based estimate of smoking pack-years: a guess at how much someone has smoked, made from their DNA chemistry and used instead of asking them. The logic is sensible on its face (smoking kills people, and people misremember their own habits, or whatever). But the material behind this piece does not describe that component, its weight in the formula, or how closely it tracks real smoking histories. I can explain why a death-trained clock would want a smoking stand-in. I can’t tell you how this one is built.&lt;&#x2F;p&gt;
&lt;p&gt;That isn’t settled here.&lt;&#x2F;p&gt;
&lt;p&gt;The second is GrimAge2. It exists, in the sense that studies report it: the HIV cohort above used both versions, and a study of organophosphate pesticide metabolites ran GrimAge alongside GrimAge2. In that pesticide study, the estimated mixture effect rose at higher exposures for Horvath, Hannum and PhenoAge but not for either GrimAge, and the effect sizes for the GrimAge clocks were modest. What the second version changed is not in my sources, so I won’t guess. (Sequels, in my experience, tend to follow originals that did well.)&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-moves-grimage-and-why-trials-keep-asking&quot;&gt;What moves GrimAge, and why trials keep asking&lt;&#x2F;h2&gt;
&lt;p&gt;Does it move? Sometimes. In adults with HIV, starting antiretroviral therapy partly reversed acceleration on Horvath, Hannum, GrimAge and PhenoAge, but GrimAge and PhenoAge were still significantly elevated compared with HIV-negative controls after two years. In the CALERIE trial, 25% calorie restriction over two years slowed DunedinPACE by about 3% but produced no statistically significant change in GrimAge. The small TRIIM study, on the other hand, reported a two-year decrease in GrimAge relative to chronological age. (GrimAge responded to one study’s attentions and not the other’s, a pattern we would describe less politely in a person.)&lt;&#x2F;p&gt;
&lt;p&gt;Each party here is acting sensibly. Researchers want a measure linked to outcomes people actually care about, and GrimAge is one. Test sellers want a number customers understand, and an age is that number. Trial designers want an endpoint that shows up in two years rather than thirty, and a blood draw does. The planning meeting more or less writes itself: “Hey, so the mortality endpoint needs a few decades and thousands of people. What if we just ran the clock?” (That quote is invented, though not much.)&lt;&#x2F;p&gt;
&lt;p&gt;The 2026 systematic review is where that meeting meets reality. It found no trial-level validation of any epigenetic clock as a surrogate endpoint, rated the certainty of evidence very low across the board, and concluded the clocks are best treated as exploratory or adjunctive markers of response. I mean, it’s fine. That is what early biomarkers are. It just means a falling GrimAge in a trial is a falling GrimAge, and nobody has yet shown it is a falling death rate.&lt;&#x2F;p&gt;
&lt;p&gt;There is also the matter of what gets measured. Almost all clocks in commercial tests are trained on blood, and the same article in The Scientist notes that GrimAge’s links to disease and death come from blood studies, adding that “this linkage has not been established for saliva-based clocks.” Anyway, the death part of the clock was learned from blood.&lt;&#x2F;p&gt;
&lt;p&gt;And the only way to check whether lowering GrimAge actually helps is to go back to the answer key and count deaths, which is the one column every trial participant signed up hoping to stay out of.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>The Horvath Clock: The Pan-Tissue Clock That Started It</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/horvath-clock/"/>
        <id>https://bioclocktest.com/horvath-clock/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/horvath-clock/">&lt;p&gt;In 2013, Steve Horvath published a formula that reads chemical tags at 353 particular spots on human DNA and guesses how old the person is. It works on almost any tissue you hand it: blood, brain, liver, and a long list of cell types besides. Its typical miss is about 3.6 years in either direction, averaged across tissues.&lt;&#x2F;p&gt;
&lt;p&gt;That formula is the Horvath clock, and it is roughly what people mean when they say “epigenetic clock” without specifying which one. It was trained to predict exactly one number, your age in years, and it is good at that. (Horvath later worked on several of the clocks that replaced it in the headlines, including PhenoAge and GrimAge, which is sort of like building the first car and then also the minivan.) What the clock’s leftover errors mean is the part that is not settled, and it is most of the interesting part.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-the-horvath-clock-is-reading&quot;&gt;What the Horvath clock is reading&lt;&#x2F;h2&gt;
&lt;p&gt;Your DNA carries small chemical tags called methyl groups, and at certain spots, called CpG sites, the share of cells carrying a tag drifts in a fairly steady way as you get older. (The site has a whole article on the chemistry, so this is the short version.) There are hundreds of thousands of candidate sites. Horvath’s clock looks at 353 of them, multiplies each reading by a weight, adds it all up, and converts the total into years.&lt;&#x2F;p&gt;
&lt;p&gt;That is it. It is a weighted sum. There is no little hourglass in there.&lt;&#x2F;p&gt;
&lt;p&gt;The clever part was the input. Horvath pooled publicly available methylation datasets covering 51 different non-cancerous tissue and cell types, so the clock had to find sites that tick at the same rate in a liver as in a white blood cell. That is why it gets called the pan-tissue or multi-tissue clock, and it is the thing that set it apart from the blood clock that arrived the same year.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-you-build-a-clock-that-works-on-51-tissues&quot;&gt;How you build a clock that works on 51 tissues&lt;&#x2F;h2&gt;
&lt;p&gt;The selection method is described by UCSF’s research program as “elastic net regression aimed at producing the most accurate estimate of chronological age.” In plain words: a computer was given the age of every sample donor, tried hundreds of thousands of DNA spots as clues, and kept the few hundred that best predicted birthdays while being penalized for keeping too many.&lt;&#x2F;p&gt;
&lt;p&gt;Think of the guess-your-age booth at a county fair. The guesser glances at gray hair, posture, whether you still call it “the Twitter,” and names a number (the booth’s error rate is not published). Elastic net is that guesser told that every extra clue costs money, so they had better pick only the ones that pay. It ended up paying for 353.&lt;&#x2F;p&gt;
&lt;p&gt;Because the training data ran across ages and tissues, the clock works on children and even prenatal samples, which is why researchers group it with the life course clocks. Clocks of this type correlate with chronological age at better than 0.90 for people between 20 and 100, which in the life sciences is the kind of number people frame.&lt;&#x2F;p&gt;
&lt;p&gt;Then there is Hannum.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;hannum-the-blood-clock-from-the-same-year&quot;&gt;Hannum, the blood clock from the same year&lt;&#x2F;h2&gt;
&lt;p&gt;Also in 2013, Hannum and colleagues published a clock of their own. It uses 71 CpG sites, it was built for blood, and it does well in blood. Its sites were picked with the same elastic net approach aimed at the same target. The Hannum model was validated in tissues beyond whole blood, and the same paper reported an accelerated ageing rate in tumor tissue, which in some sense was an early hint that these numbers might carry information beyond the calendar.&lt;&#x2F;p&gt;
&lt;p&gt;The two arrived together and have stayed together, like the names on a law firm door. When the DunedinPACE paper wanted to show how it compared with existing clocks, it lined itself up against Horvath, Hannum, PhenoAge and GrimAge. A 2026 systematic review of epigenetic ageing in cancer patients found the Levine (that is, PhenoAge), Horvath and Hannum clocks were the most common methods used. Researchers tend to run Horvath and Hannum side by side, presumably because once the chip has been read, computing one more weighted sum is not much work.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-it-means-when-the-horvath-clock-says-you-are-older&quot;&gt;What it means when the Horvath clock says you are older&lt;&#x2F;h2&gt;
&lt;p&gt;If the clock says you are 52 and your passport says 47 (a made-up person), the five-year gap is called age acceleration. There is a refined version, intrinsic epigenetic age acceleration, which also adjusts for blood cell counts, so that a change in which cells happen to be in the tube does not get counted as ageing. It is a bit like splitting a dinner bill evenly and then checking who owes extra: the leftover is where the information is, but it also contains whoever forgot the tip.&lt;&#x2F;p&gt;
&lt;p&gt;The research gives these gaps some weight. First-generation clocks put people living with HIV about 5 to 7 years older than HIV-negative people of the same age, though that may be partly the immune cell shifts again. Starting antiretroviral therapy partly reversed the acceleration over two years in a substudy of the NEAT001&#x2F;ANRS143 trial. In the NHANES cohort, followed for 17.5 years, Horvath acceleration predicted overall mortality with a hazard ratio of 1.12 per five years, and predicted cancer mortality specifically.&lt;&#x2F;p&gt;
&lt;p&gt;So what is the gap? There are, roughly, three ways to read this.&lt;&#x2F;p&gt;
&lt;p&gt;The first is that it is mostly noise. Elysium Health reported that technical replicates of published clocks, including Horvath’s, came back three to nine years apart. Same sample, run twice. To be fair, Elysium sells its own test built a different way, so it has a reason to say so, but having a reason to say something does not make it false. Somewhere a technician is saying, “Hey, can you rerun tube 14? It aged four years over lunch.”&lt;&#x2F;p&gt;
&lt;p&gt;The second is that the gap is real biology, specifically the biology of ordinary ageing. A summary from CD Genomics says the purely chronological training goal makes these clocks “better at capturing normal aging than pathological aging related to disease and mortality risk.” Look, normal ageing is a real thing to measure, and a hazard ratio of 1.12 is not nothing. It is just smaller than GrimAge’s 1.50 in the same cohort.&lt;&#x2F;p&gt;
&lt;p&gt;The third reading is the uncomfortable one. The clock was trained to make the gap as small as possible. Whenever a DNA spot tracked something other than birthdays (smoking, illness, a bad decade) in a way that spoiled the birthday guess, the method had a reason to drop it. So the gap is the stuff the clock was built to ignore and did not quite manage to.&lt;&#x2F;p&gt;
&lt;p&gt;Push that to its end. Here is an invented scenario: someone builds a perfect version, error zero. It reads 353 spots on your DNA and returns your exact birth date. Every customer has an age acceleration of zero. Everyone is exactly on schedule. It is a very expensive way to read a driver’s license, and it is where the training objective points. (I would rather not think about what a marketing department has planned for a test that tells everyone they are precisely average.)&lt;&#x2F;p&gt;
&lt;p&gt;The data fits this reading uncomfortably well. In the Add Health study of younger U.S. adults, sociodemographic and lifestyle factors were associated with ageing more often in clocks trained on current or changing health than in clocks trained on age alone, such as Horvath. In another study, living in lower-income census tracts was linked to older PhenoAge and GrimAge, but not Horvath or Hannum. Between us, a clock that cannot see your neighborhood is not the one anybody would lead a brochure with.&lt;&#x2F;p&gt;
&lt;p&gt;Which reading is right? It is not settled. The three are not exclusive, and I assume most gaps are some of each.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;where-the-first-clock-still-gets-used&quot;&gt;Where the first clock still gets used&lt;&#x2F;h2&gt;
&lt;p&gt;The original did not retire. It got a successor of its own in 2018, the skin and blood clock, which does better on keratinocytes, cheek cells, fibroblasts, saliva and blood. The 2013 version still shows up in the standard battery: the TRIIM-X trial uses it alongside Hannum, PhenoAge and GrimAge, a lung cancer survival study used it with three others, and a Parkinson’s analysis in the PPMI cohort found that in men, acceleration across Horvath v1, PhenoAge and GrimAge went with worse tremor, gait and rigidity. First-generation clocks are considered well suited to tissue-specific ageing and to benchmarking lab platforms, which is a polite way of saying they are the ruler you check the new rulers against.&lt;&#x2F;p&gt;
&lt;p&gt;They do get beaten. A 2022 study built clocks with newer feature-selection methods that outperformed Hannum, Horvath and Weidner on two outside datasets. That is the normal fate of a first version, and the field seems fine with it. (The citation count presumably softens the blow.)&lt;&#x2F;p&gt;
&lt;p&gt;The first clock was built to know your birthday. Everything else it tells you is the part it got wrong.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>&#x27;Methylation Test&#x27; Means Three Different Things</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/methylation-test/"/>
        <id>https://bioclocktest.com/methylation-test/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/methylation-test/">&lt;p&gt;I spent part of this week trying to pin down what a methylation test is, and the honest answer is that it depends on who is selling it to you. The phrase gets used for at least three different products: an epigenetic age test, which reads chemical tags on your DNA and turns them into a “biological age”; an MTHFR test, which looks at a gene; and a homocysteine test, which measures a substance in your blood. They share a word because all three sit somewhere near the body’s methylation chemistry. They share very little else.&lt;&#x2F;p&gt;
&lt;p&gt;The first kind is what this site covers. It reads which spots on your DNA carry a methyl group (a tiny cluster of one carbon and three hydrogens) and feeds that pattern into a formula trained to guess age. The MTHFR and homocysteine kinds ask about the gene and blood chemistry side of methylation, which is a different question from how old your DNA tags look. The sources behind this piece are about the clocks. On MTHFR and homocysteine they say nothing, so I am not going to put numbers on them. The point is that one search term covers a gene, a blood level and an age estimate, and the useful thing to know first is which one you are looking at.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-one-word-covers-a-gene-a-blood-level-and-a-clock&quot;&gt;Why one word covers a gene, a blood level and a clock&lt;&#x2F;h2&gt;
&lt;p&gt;There is a neat illustration of the gap in the CALERIE trial’s data. The CALERIE Genomic Data Resource lists “whole-genome single-nucleotide polymorphism genotypes” in one bucket and “three-timepoint-longitudinal DNA methylation” in another. Genotype is what your DNA says. Methylation is what is stuck to it, and it can be measured at three timepoints because it changes.&lt;&#x2F;p&gt;
&lt;p&gt;A gene test is basically a test of the first bucket. An age clock is a test of the second. A blood level is in neither bucket (it is chemistry floating around your veins). In ordinary speech all three become “methylation,” the way “football” means one sport in Ohio and a different one in London, and both crowds are completely sure they ordered correctly.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;who-gets-paid-when-methylation-test-stays-vague&quot;&gt;Who gets paid when “methylation test” stays vague&lt;&#x2F;h2&gt;
&lt;p&gt;This is the part where I follow the money, one link at a time, because that is the only skill I really have.&lt;&#x2F;p&gt;
&lt;p&gt;Link one is the lab. A lab gets paid to produce methylation readings, and it can do that in different ways. According to The Scientist, some providers use Illumina arrays and others use whole-genome bisulfite sequencing, and “these are fundamentally different measurement methods.” The lab’s incentive is to read a lot of sites cheaply. It is not to make its numbers match the lab down the street (which, to be fair, is not trying to match it either).&lt;&#x2F;p&gt;
&lt;p&gt;Link two is the clock. Researchers get paid (in grants, citations and occasionally licenses) to build formulas that predict age from those readings. The Horvath pan-tissue clock uses 353 CpG loci and lands within about 3.6 years of chronological age on average across tissue types. I would like to stop on the 353 for a moment. Not 350, not a round 400. Three hundred and fifty-three specific spots out of a genome’s worth of options, chosen by a statistical method that kept exactly the ones that helped and dropped the rest, and from those spots you can guess a stranger’s age within a few years from a piece of tissue. That is strange and good. Almost all the clocks used in commercial tests, The Scientist reports, are trained on blood.&lt;&#x2F;p&gt;
&lt;p&gt;Link three is the company that sells you the kit. It gets paid per kit, and the kit has to sound like it measures something important. So you get numbers like these: Elysium Health says its Index test examines “more than 100,000 CpG sites” on a custom chip, and TruDiagnostic says it measures over 1 million DNA sites per test. (Horvath, you will recall, made do with 353.) Meanwhile most companies selling these kits do not say which clock they use. More sites is a feature you can put on a box. Which formula is a footnote.&lt;&#x2F;p&gt;
&lt;p&gt;“Can we just call it a methylation test?” a product manager says, in a meeting I am inventing. “People search for that.”&lt;&#x2F;p&gt;
&lt;p&gt;They do.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-the-clock-version-of-a-methylation-test-actually-measures-and-how-well&quot;&gt;What the clock version of a methylation test actually measures, and how well&lt;&#x2F;h2&gt;
&lt;p&gt;Here is where the vagueness starts to cost something, because the clock tests have real measurement problems that the gene and blood versions do not share. As of 2026, according to The Scientist, the field has no reference standards, meaning you cannot send one sample to two validated labs and get the same biological age back. That is a small sentence describing a fairly large hole.&lt;&#x2F;p&gt;
&lt;p&gt;Even within one lab, the numbers wander. Elysium Health, in its own write-up, says technical replicates of the published Horvath, Hannum and PhenoAge clocks produced deviations of three to nine years between measurements of the same sample. It says its own test agrees across replicates within zero to 1.5 years. (A company reporting that its competitors are noisy and it is not is a genre, and I would read it as one.) DunedinPACE took a different route: its builders threw out probes with poor test-retest reliability before training, and eLife describes the result as having “high test-retest reliability,” above 0.90 by its developers’ count.&lt;&#x2F;p&gt;
&lt;p&gt;Then there is the sample itself. A blood-trained clock applied to saliva without correction can produce large errors, because saliva’s cell types carry different methylation signatures. So a cheek swab read by a blood formula is, i don’t know, a bit like a bathroom scale calibrated for people being used on luggage. It gives a number. The number is in pounds.&lt;&#x2F;p&gt;
&lt;p&gt;None of this makes the clocks useless. The AMA Journal of Ethics, in a December 2025 commentary, says the tests have been shown to be fairly reliable indicators of risk for certain age-related diseases. It also says they cannot tell anyone how old they “really” are and calls biological age a fiction. Both of those can be true. It’s fine.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;where-the-chain-ends-up&quot;&gt;Where the chain ends up&lt;&#x2F;h2&gt;
&lt;p&gt;Link four is the supplement. If you sell the test, you know the customer’s age result, and if you also sell products, the result is a very warm lead. The AMA Journal of Ethics commentary describes a healthy 56-year-old, AA, whose direct-to-consumer result said he was “biologically 65-years-old.” He then followed the manufacturer’s recommendations and began courses of the company’s brain health and cellular repair supplements. It is always nice for a company when its diagnostic happens to find the exact problem its product line was designed for, and I am sure everyone involved was as surprised as we are.&lt;&#x2F;p&gt;
&lt;p&gt;The Stanford researcher Tony Wyss-Coray has called this market “a bit like the wild West,” noting that many tests come bundled with products never tested in a human clinical trial. That is a fair description of a market where the measurement has no reference standard and the seller of the measurement is also the seller of the fix.&lt;&#x2F;p&gt;
&lt;p&gt;Link five is the search engine, which gets paid when people click, and does not much care which of three meanings they had in mind. Here is a scenario I am making up. Someone’s relative mentions MTHFR at dinner. She searches the phrase, clicks the first ad, and orders what turns out to be an epigenetic age kit. It reports that she is seven years older than her birthday. Unsettled, she orders a second kit from another company, which uses a different clock on a different platform and reports that she is two years younger. She now has two biological ages, a nine-year spread, a supplement subscription, and still no information about the gene she originally asked about. Somewhere a third company is preparing a mailer about homocysteine. Who exactly was she supposed to ask?&lt;&#x2F;p&gt;
&lt;p&gt;I write about this for a living, which mostly means reading the same methods section until it stops looking like a hostage note, and even I had to check which test was which. That is not her failure. Every link in the chain was paid to make the word sound more important and nobody was paid to make it more specific. The AMA Journal of Ethics asked companies to be clearer about “how the tests work, how reliable they are, and what they can and cannot show,” and also warned that results can make age discrimination look justified. That second worry gets less attention, lol, given that it involves employers rather than customers.&lt;&#x2F;p&gt;
&lt;p&gt;Which is how a single ambiguous phrase ends up doing the job no lab has managed yet. The labs cannot agree on your age, but the word “methylation” reliably gets all three industries the same customer.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>PhenoAge: A Clock Trained on Blood Tests</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/phenoage/"/>
        <id>https://bioclocktest.com/phenoage/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/phenoage/">&lt;p&gt;Here is a small fact about PhenoAge: the expensive version, the one read off your DNA, is trained to predict the cheap version, the one read off an ordinary blood test. The cheap version, from a PLOS Medicine study built on NHANES data, is “calculated as a linear combination of chronological age and 9 multi-system clinical chemistry biomarkers.” Those include creatinine, blood glucose, C-reactive protein and lymphocyte percentage. You take a person’s age and nine numbers a routine lab already reports, multiply each by a weight, add them up, and out comes an age.&lt;&#x2F;p&gt;
&lt;p&gt;The expensive version, called DNAm, reads methylation at hundreds of CpG sites on the genome. Its target is not the person’s birthday. It is that nine-marker blood score. So the short answer is that the clock is a weighted blend of blood chemistry, and its molecular sibling is an estimate of that blend made from a different kind of sample. Both are meant to say something about disease and death beyond what age alone says. How precisely either one describes a single person is less settled.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-phenoage-is-built-from-the-dumb-version&quot;&gt;What PhenoAge is built from, the dumb version&lt;&#x2F;h2&gt;
&lt;p&gt;I mean “dumb” as a compliment here. The blood version requires no special equipment. The inputs are things a hospital lab produces every day for other reasons (kidney function, sugar, inflammation, white cell mix, etc.), and the output is arithmetic.&lt;&#x2F;p&gt;
&lt;p&gt;Stated with the gravity it deserves: whereas the subject has had blood drawn; and whereas the laboratory has, in the ordinary course of its business, reported a creatinine value; and whereas the parties agree that creatinine shall be multiplied by a coefficient; now therefore the subject is 47. No sequencing, no kit and no shipping label are involved, and the coefficients fit on an index card.&lt;&#x2F;p&gt;
&lt;p&gt;The dumb version is also hard to improve on by much. A cancer cohort study built GeroAge, which uses only six blood markers plus age, and reported R=0.99 agreement with PhenoAge and similar power to predict overall survival in NHANES data. Going the other direction, a 25-biomarker clock trained on UK Biobank data (n = 306,116) with an Elastic-Net Cox model, picking from 60 circulating biomarkers, reached a C-Index of 0.778 for mortality. That was a lift of 0.028 over PhenoAge, which the write-up describes as an 11% relative increase in predictive value. Drop three markers and nothing much happens. Add sixteen and you get a real but modest bump.&lt;&#x2F;p&gt;
&lt;p&gt;I feel like that is the most informative fact about the blood clock. Most of the signal lives in a handful of common lab values.&lt;&#x2F;p&gt;
&lt;p&gt;It is a bit like a quarterback passer rating. Nobody thinks the formula is the essence of quarterbacking, but it is built from numbers that already exist, and it tracks winning well enough that people keep using it.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-dnam-phenoage-reaches-the-same-number-by-a-longer-road&quot;&gt;How DNAm PhenoAge reaches the same number by a longer road&lt;&#x2F;h2&gt;
&lt;p&gt;The first epigenetic clocks, such as Horvath’s, were trained to predict chronological age, which is a strange thing to predict (you can just ask). CD Genomics describes what happened next: second-generation clocks “changed the training objective from predicting chronological age to predicting clinical phenotypes, incidence of disease, and all-cause mortality associated with aging.” DNAm PhenoAge is one of those. The DunedinPACE paper cites it as Levine et al. 2018. The Scientist lists Steve Horvath as having developed and validated it, alongside the Horvath clock, GrimAge and DunedinPACE.&lt;&#x2F;p&gt;
&lt;p&gt;The sophisticated road goes like this. You draw blood, or collect some other sample. You extract DNA. You run it on a methylation array (one adolescent lead-exposure study, for instance, used Infinium MethylationEPIC BeadChips). You read methylation at hundreds of CpG sites. You feed those through a regression whose weights were chosen so the result lines up with phenotypic age, the nine-marker blood score. And you arrive at an estimate of a number the lab could have handed you from the same tube.&lt;&#x2F;p&gt;
&lt;p&gt;Here is the conversation I imagine at some point in the pipeline:&lt;&#x2F;p&gt;
&lt;p&gt;“Hey, we already have the glucose.”&lt;&#x2F;p&gt;
&lt;p&gt;“Sure, but this is the glucose’s methylation.”&lt;&#x2F;p&gt;
&lt;p&gt;That is unfair, but only a little. DNAm PhenoAge is what powers at least one consumer product: Elysium Health says its Index test uses the DNAm PhenoAge biomarker as the basis of its biological age measurement. The methylation version can be sold as a kit, because the sample can be collected outside a clinic. The blood version needs a lab order and a phlebotomist. In product terms, the methylation clock is the one that gets to flirt with customers directly, while the blood panel stays home with the doctor.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-train-a-clock-to-predict-another-clock&quot;&gt;Why train a clock to predict another clock&lt;&#x2F;h2&gt;
&lt;p&gt;One reading is that methylation captures something the blood panel misses. Chemistry values bounce around with last night’s dinner or a cold; a methylation pattern might reflect a slower, accumulated state. If so, the long road is worth walking, because you end up near the same place with less day-to-day weather in the measurement.&lt;&#x2F;p&gt;
&lt;p&gt;The second reading is that the methylation version is mostly a costlier, noisier reproduction of a number you could compute directly, and that its appeal is the sample, not the science. The facts here do not choose between these. The clock was built to agree with the blood score, so agreement alone cannot tell you which reading is right.&lt;&#x2F;p&gt;
&lt;p&gt;There is some evidence the second-generation design picks up life circumstances better than first-generation clocks. In the Add Health study of younger U.S. adults, sociodemographic and lifestyle factors “were more often associated with biological aging” in clocks like PhenoAge, GrimAge and DunedinPACE than in clocks trained on chronological age alone, like Horvath. That’s fine as far as it goes. It says the target matters. It does not say the methylation route beats the blood route to the same target.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-much-the-number-moves-when-you-measure-it-twice&quot;&gt;How much the number moves when you measure it twice&lt;&#x2F;h2&gt;
&lt;p&gt;This is where the long road gets weird. Elysium Health, which sells a test based on DNAm PhenoAge, also reports that technical replicates of published clocks (Horvath, Hannum, PhenoAge) produced deviations of three to nine years between replicate measurements. A replicate is the same sample run again. The person did not age between runs; the lab work just disagreed with itself.&lt;&#x2F;p&gt;
&lt;p&gt;Sample type adds its own problem. The Scientist notes that almost all clocks used in commercial tests are trained on blood, and that applying a blood-trained clock to saliva “without correction can produce substantial errors,” because the cells in saliva carry different methylation signatures.&lt;&#x2F;p&gt;
&lt;p&gt;Now a scenario, which is invented. Suppose someone runs the same sample twice and gets 41 and then 49. They average to 45. To tighten it they run it four more times, then sixteen, until the standard error is smaller than a birthday. At that point they have spent a great deal of effort to estimate, with some precision, a number that a creatinine test, a glucose test and seven friends would have produced in an afternoon. Then they get a cold, the blood panel shifts, and the thing they were so carefully estimating has moved anyway.&lt;&#x2F;p&gt;
&lt;p&gt;(I write about blood tests on the internet, so this is roughly the most exciting plot my job allows.)&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-changes-phenoage-according-to-the-studies&quot;&gt;What changes PhenoAge, according to the studies&lt;&#x2F;h2&gt;
&lt;p&gt;Studies use PhenoAge constantly, usually alongside other clocks, and the clocks do not always move together. People with HIV who started antiretroviral therapy saw epigenetic age acceleration partly reverse across Horvath, Hannum, GrimAge and PhenoAge, but PhenoAge and GrimAge stayed significantly elevated versus HIV-negative controls after two years. In an organophosphate pesticide study, estimated mixture effects rose at higher exposure for Horvath, Hannum and PhenoAge but not GrimAge or GrimAge2. A cortisol-to-DHEAS ratio correlated with Hannum, Horvath Skin &amp;amp; Blood and PhenoAge, but not Horvath or GrimAge. Among men with Parkinson’s disease in the PPMI cohort, faster aging on Horvath v1, PhenoAge and GrimAge went with worse tremor, gait and rigidity.&lt;&#x2F;p&gt;
&lt;p&gt;So it moves with illness, treatment, exposures and hormones, I guess, in ways that partly overlap with its siblings. It is also not the strongest blood-derived predictor around: a UK Biobank proteomic clock built from 2,923 proteins in 53,021 people showed stronger age-adjusted links to mortality and 18 incident diseases than PhenoAge did.&lt;&#x2F;p&gt;
&lt;p&gt;Whether moving the number means anything clinically is open. A 2026 systematic review concluded that epigenetic clocks should currently be treated as exploratory or adjunctive markers of intervention response, not validated surrogate endpoints. A commentary in the AMA Journal of Ethics goes further, arguing these tests cannot tell how old a person “really” is and calling biological age a fiction. I suppose a fiction can still be a useful one. Passer rating is too.&lt;&#x2F;p&gt;
&lt;p&gt;The logical end of all this is a model that predicts DNAm PhenoAge from creatinine, glucose, C-reactive protein and six other routine lab values. It will agree with DNAm PhenoAge extremely well, and it will be the original.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>Proteomic Clocks and Organ Ages</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/proteomic-aging-clocks/"/>
        <id>https://bioclocktest.com/proteomic-aging-clocks/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/proteomic-aging-clocks/">&lt;p&gt;Someone measured 2,923 proteins in the blood of 53,021 people, in something called the UK Biobank Pharma Proteomics Project, and then built a model to guess how old each of those people was. The researchers already knew how old each of those people was. (It was in the spreadsheet, next to their participant IDs or whatever.) This is not a mistake. It is basically the entire method, and it works.&lt;&#x2F;p&gt;
&lt;p&gt;A proteomic aging clock is a statistical model that reads the levels of many proteins in a blood plasma sample and turns them into one number shaped like an age. The useful part is the gap between that number and your actual age. A positive gap is read as faster aging, and in the big biobank studies that gap tracks death and disease better than some older measures. The fair description of the field is early: the best studies are large and replicate across countries, but the clocks run on different machines, in different populations, with different math. The newest offshoot, organ-specific ages, is mostly a promising idea that occasionally has a price attached.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-a-proteomic-aging-clock-reads-from-your-blood&quot;&gt;What a proteomic aging clock reads from your blood&lt;&#x2F;h2&gt;
&lt;p&gt;DNA methylation clocks, which the rest of this site covers at length, read chemical tags on DNA. Those tags are a bit like the settings on a machine. Proteins are closer to the output: the stuff cells actually make and ship. A review titled “Proteomic ageing clocks: methods, mechanisms and the muscle connection” says the clocks “are predominantly built from blood plasma, capturing inflammatory, metabolic, and extracellular proteins including secreted signaling factors.”&lt;&#x2F;p&gt;
&lt;p&gt;Plasma is used because plasma is what you can get. Biopsying 53,021 livers is not on the table. (The volunteers, reasonably, would have had opinions.) So the clocks read whatever happens to be floating in the blood, including proteins made in particular organs that leaked out or were sent out. The same review notes that muscle clocks rely on “organ-enriched circulating proteins rather than biopsy-derived data.” I am fond of “organ-enriched circulating proteins.” These are proteins that have left home, are out circulating, and still carry traces of where they spent the night.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-the-first-thing-the-clock-predicts-is-your-birthday&quot;&gt;Why the first thing the clock predicts is your birthday&lt;&#x2F;h2&gt;
&lt;p&gt;Here is the problem every clock builder faces. You have a giant table: tens of thousands of people down the side, thousands of proteins across the top. To train a model you need a target column, the thing it learns to predict. The column you actually want, “how worn out is this body,” does not exist. Nobody has a ruler for that. What does exist, for every participant, filled in correctly, is chronological age.&lt;&#x2F;p&gt;
&lt;p&gt;I imagine the meeting goes something like this (I made it up): “We have 2,923 proteins and fifty-three thousand people, what do we train on?” “Age. It’s the only column nobody left blank.”&lt;&#x2F;p&gt;
&lt;p&gt;Let the record therefore reflect that the principal output of the instrument is an estimate of a quantity supplied to the instrument as an input, and that the instrument’s accuracy is assessed by the fidelity with which it returns that input to its owners. The study published under PubMed ID 39117878 trained on 45,441 UK Biobank participants, settled on 204 proteins, and reported a Pearson correlation of 0.94 with chronological age. It held up elsewhere: 0.92 in a Chinese biobank cohort of 3,977 people and 0.94 in a Finnish cohort of 1,990. That is good. The protein signature of getting older is not a quirk of British blood.&lt;&#x2F;p&gt;
&lt;p&gt;But a perfect age predictor would be worthless. If the model hit your birthday every time, it would tell you nothing beyond your birthday. All the value lives in the misses. So the incentive is strange: you want a model very good at guessing age, but not too good, and the residual is the product.&lt;&#x2F;p&gt;
&lt;p&gt;Consider a made-up startup that estimates age by counting candles on birthday cakes in your phone’s photo roll. Its correlation with chronological age would be close to 1. It would beat every protein clock ever published. Its “candle age acceleration” would mostly measure whether your friends ran out of candles at 40 and switched to the number-shaped kind, and people with that acceleration would be flagged as aging slowly. It would raise money on the correlation alone. Then it would launch organ ages, with the heart read off whichever candle sits nearest the frosting rose.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;who-pays-for-the-gap-and-what-it-buys&quot;&gt;Who pays for the gap, and what it buys&lt;&#x2F;h2&gt;
&lt;p&gt;The next link in the chain. Matching birthdays proves the method; showing that the gap predicts something bad is what gets attention, and renewals. The PAC paper reports that its clock “showed stronger age-adjusted associations with all-cause mortality and 18 incident diseases than PhenoAge, BioAge, or leukocyte telomere length.” The 39117878 study tied proteomic aging to telomere length, to a frailty index, and to reaction time, which covers the biological, physical and cognitive bases in one go.&lt;&#x2F;p&gt;
&lt;p&gt;The proteins doing the work look like aging biology, which is reassuring. In PAC they were enriched in immunoinflammatory responses, cellular senescence, extracellular matrix remodeling and vascular biology. In the other clock they were involved in hormone regulation and reproduction, neuronal structure and function, and development and differentiation. The funders on the 39117878 paper include the Wellcome Trust, the British Heart Foundation (both listed with the same grant number, CH&#x2F;1996001&#x2F;9454, for reasons the facts don’t explain), and NIH grant P30 DK043351. A heart charity is going to want disease prediction. (Few charities exist to confirm birthdays.) It’s fine. That is what the gap is for.&lt;&#x2F;p&gt;
&lt;p&gt;Now a puzzle. PAC’s correlation with chronological age was 0.77 (Spearman), noticeably below the 0.94 elsewhere. One reading: PAC gave up some birthday accuracy and got more biology in return, since the misses are where mortality lives. The other reading: these are different models, different correlation measures and different choices, and the two numbers were never built to be lined up side by side. Both fit what has been published.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-a-blood-test-becomes-a-liver-age&quot;&gt;How a blood test becomes a liver age&lt;&#x2F;h2&gt;
&lt;p&gt;The chain keeps going. If one clock is good, a clock per organ sounds better. The recipe is basically to take proteins enriched in a given organ and build a clock from just those. The study “Multimodal clocks of human aging” built a “three-tiered aging framework comprising the CC-clock, MM-clock, and organ-associated aging clocks,” and argued that plasma protein clocks “serve as efficient proxies for systemic physiological capacity.” (What exactly CC and MM stand for is not in the facts I have.) It also flagged age-dependent buildup of coagulation factors as a driver of multi-organ senescence and inflammation, which is a causal claim, and causal claims in this field are the expensive kind.&lt;&#x2F;p&gt;
&lt;p&gt;Note what “organ age” means here. It is not a scan of the organ. It is proteins from that organ, observed after they have traveled to the arm, run through a model whose training target was mostly your birthday. Meanwhile, on the consumer side, TruDiagnostic’s TruAge says it covers “organ-specific aging for 11 body systems including brain, heart, and liver,” across 64 biomarkers; how those organ numbers are built is not something the facts here settle, and another article on this site covers that product.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-reliable-the-proteomic-aging-clock-is-right-now&quot;&gt;How reliable the proteomic aging clock is right now&lt;&#x2F;h2&gt;
&lt;p&gt;The big biobank numbers are strong. The smaller studies show how fragile the setup can be. A Moscow study enrolled 350 adults aged 19.8 to 99.1, measured plasma by mass spectrometry, and got a mean absolute error of 11.83 years from an 8-protein model. The state-fair guy who guesses your age for a dollar looks at your face and probably does better, lol. The authors also found that age and sex distributions “differed strongly across 27 MS batches,” which they called “a major confound.” Put plainly, the machine runs may have known who was old before the proteins did.&lt;&#x2F;p&gt;
&lt;p&gt;The review “Proteomic aging clocks in epidemiological studies” says the clocks “vary widely in their assay platforms, study populations, and modeling strategies,” which may limit generalizability. It also lists three possible jobs: discovery biomarkers, predictive biomarkers and response biomarkers. That last one, a number that moves when an intervention works, is the one everybody wants, and it is the one nobody has shown yet. An older paper, on the other hand, argued that methylation clocks have “unmatched accuracy” against other omics. That contest is not settled either.&lt;&#x2F;p&gt;
&lt;p&gt;You and I both know which figure ends up on the slide. Anyway, the proteins in your blood know how old you are. That part is done.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>Tally Health&#x27;s Cheek-Swab Age Test</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/tally-health/"/>
        <id>https://bioclocktest.com/tally-health/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/tally-health/">&lt;p&gt;The way a cheek-swab age test works is that you rub a swab along the inside of your cheek, mail it off, and a lab reads chemical tags on the DNA in those cheek cells and feeds them into a model that turns the tags into a number of years. That is the product Tally Health sells: a swab, and a clock called CheekAge that reads it. The number is supposed to be your biological age, as opposed to the one on your driver’s license.&lt;&#x2F;p&gt;
&lt;p&gt;The plain answer to what is unusual here is the sample. The Scientist, surveying the consumer market, put it this way: “Almost all epigenetic clocks used in commercial age prediction tests are trained on blood samples.” A cheek is not blood. So the real question is whether a clock built for cheek cells measures aging as well as the blood clocks do, and the honest answer from the material this site works from is that it does not say. There are no published accuracy figures for CheekAge in the facts behind this article, so how well it performs is not settled here (which is a strange thing to admit in the second paragraph, but there it is).&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-the-tissue-matters-at-all&quot;&gt;Why the tissue matters at all&lt;&#x2F;h2&gt;
&lt;p&gt;An epigenetic clock is basically a weighted list of spots on the genome, called CpG sites, where methylation tends to drift with age. You train the model on samples from people whose ages you know, it learns which spots to weight, and then it guesses ages for new people. The details are on the site’s article about how clocks are built. The point is that the model learns from whatever tissue it was trained on, and tissues are not interchangeable.&lt;&#x2F;p&gt;
&lt;p&gt;The Hannum clock, for instance, uses 71 CpG sites and was developed on blood, where it achieves an average absolute error as low as 3 to 4 years. The Horvath clock, published in 2013, is the famous exception: a multi-tissue clock described as working across blood, brain, liver and various cell types. It uses 353 CpG loci. Sit with that for a second. Out of a genome with an enormous number of possible methylation sites, 353 of them, weighted correctly, will tell you roughly how old a piece of liver is to within about 3.6 years. That is a weird fact about biology.&lt;&#x2F;p&gt;
&lt;p&gt;It also gets messy. Blood is a mixture of cell types, and the mixture itself shifts as people age. Researchers built adjusted measures to deal with this: one version of the Horvath number regresses out blood cell counts to get at “intrinsic” aging, and one version of the Hannum number deliberately upweights blood cell composition to capture immune aging. So in blood, the same cell mix gets thrown out by one number and is the whole point of another. A cheek swab has its own cell mix. The facts here do not describe it.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;three-ways-to-read-a-cheek-swab-clock&quot;&gt;Three ways to read a cheek-swab clock&lt;&#x2F;h2&gt;
&lt;p&gt;There are, roughly, three ways to read a test like this.&lt;&#x2F;p&gt;
&lt;p&gt;The first reading is that it’s fine. Horvath showed in 2013 that age leaves a methylation signature across many tissues, so there is no law of nature saying the cheek is mute. If you train a clock on cheek cells from people of known age, it should learn the cheek’s version of the signature, and a cheek swab is a lot less unpleasant to collect than a blood draw (nobody has ever fainted at the sight of a cheek). On this reading, CheekAge is just a clock trained on a more convenient sample.&lt;&#x2F;p&gt;
&lt;p&gt;The second reading is that it’s a different instrument, and its numbers live in their own universe. Most of the research tying clocks to health outcomes was done in blood. The Add Health study, which found lifestyle and sociodemographic factors more often associated with clocks like PhenoAge, GrimAge and DunedinPACE than with Horvath, used blood from 4,237 participants aged 33 to 44. A cheek clock does not automatically inherit any of that. Comparing a cheek age to a blood age is like comparing a quarterback rating in the NFL to one in a flag football league in a parking lot. Same formula, maybe. Different game.&lt;&#x2F;p&gt;
&lt;p&gt;Here is a scenario, made up for illustration. Someone swabs a cheek and gets 41. The same week a finger-prick blood kit says 47. Which one is right? Neither company is obliged to agree with the other, so a third kit is ordered to break the tie, which says 44, and now there are three ages and an average, and the average is itself a sort of fourth clock with no validation at all. Push it further and you start swabbing different parts of the mouth. Eventually the person has a spreadsheet. I mean, it would be a very well-organized midlife.&lt;&#x2F;p&gt;
&lt;p&gt;The third reading is the least comfortable one, which is that the cheek is not really the problem. The problem is that nobody has a stable answer in blood either. The same piece in The Scientist reported that as of 2026 the field “has not yet achieved reference standards,” meaning you cannot send one sample to different validated labs and get the same readout. It also found most companies do not even say which clock they use (a choice that does spare everyone a lot of awkward follow-up questions). Against that background, worrying whether a cheek clock matches the blood clocks is a bit like worrying whether your watch matches the other watches, in a town where the watches do not match each other.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-much-a-single-sample-can-wobble&quot;&gt;How much a single sample can wobble&lt;&#x2F;h2&gt;
&lt;p&gt;The most striking number on this comes from a company selling a competing test. Elysium Health, which markets a blood-and-saliva-free-of-cheek product called Index (it uses the PhenoAge biomarker, on a custom chip covering more than 100,000 CpG sites), states that “a single epigenetic age sample processed repeatedly can result in biological age estimates that vary up to nine years.” Technical replicates of the Horvath, Hannum and PhenoAge clocks, it says, deviated by three to nine years.&lt;&#x2F;p&gt;
&lt;p&gt;Nine years is a modest amount of disagreement for one tube.&lt;&#x2F;p&gt;
&lt;p&gt;You can picture the lab: “Hey, can you rerun this one? It said 38 this morning and now it says 46.” Elysium’s pitch is that its own version shows “agreement across technical replicates within zero to 1.5 years,” which is the dull lab way of saying the sample tells the same story no matter who asks, a quality also prized in witnesses and in anyone who has been out late. Whether that claim holds up is a company describing its own product, and the facts here do not include outside checks. But the general point stands. Repeatability is a property you have to demonstrate for a specific test in a specific tissue, and the evidence for CheekAge on that front is not in front of us.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-a-tally-health-result-does-and-does-not-establish&quot;&gt;What a Tally Health result does and does not establish&lt;&#x2F;h2&gt;
&lt;p&gt;So what is the evidence a reader could want? Roughly: does the test give the same answer twice, does it track chronological age, and does it predict anything that matters, like disease or death. Blood clocks have some of this. PhenoAge’s link to all-cause mortality held up across age, race and ethnicity, education, disease count and health behaviors. The AMA Journal of Ethics, in a December 2025 commentary, said epigenetic clock tests have been shown to be fairly reliable indicators of risk for certain age-related diseases, while also calling “biological age” itself a fiction, in the sense that no test tells you how old you really are. Both of those can be true. A thermometer reading is useful without being your soul’s temperature.&lt;&#x2F;p&gt;
&lt;p&gt;The same commentary described a patient, AA, who after an epigenetic test result began courses of the manufacturer’s own brain health and cellular repair supplements, on the manufacturer’s recommendation. That is a reasonable thing for a person to do when a company hands them a number and a suggestion. It is also the incentive in miniature: the firm that measures the problem also sells the fix. Tony Wyss-Coray, a researcher quoted on the market, called it “a bit like the wild West,” with many tests coupled to products never tested in a clinical trial. The commentary’s authors also worried the numbers could cause psychological harm and make age discrimination look justified, lol, imagine an employer asking for your cheek age.&lt;&#x2F;p&gt;
&lt;p&gt;(I write about aging clocks for a living, and I have read enough replicate tables this year to believe my own clock has advanced considerably, though I have no idea which tissue to blame.)&lt;&#x2F;p&gt;
&lt;h2 id=&quot;where-this-leaves-the-cheek&quot;&gt;Where this leaves the cheek&lt;&#x2F;h2&gt;
&lt;p&gt;I don’t know whether CheekAge is a good clock. Nothing here says it is bad, and nothing here says it is good; the published validation that would decide it is not among the facts. What is settled is narrower: clocks are tissue-specific, most commercial ones are trained on blood, the same sample can come back years apart on some published clocks, and the field has no reference standard that would let any two companies, cheek or blood or whatever, agree on what a correct answer is.&lt;&#x2F;p&gt;
&lt;p&gt;Which means the most accurate statement anyone can currently make about a cheek-swab age is that it is exactly the age of your cheek, according to your cheek.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>Telomeres: The Ageing Marker That Underperformed</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/telomeres/"/>
        <id>https://bioclocktest.com/telomeres/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/telomeres/">&lt;p&gt;A large study in the UK Biobank Pharma Proteomics Project (53,021 people, 2,923 proteins measured in their blood) built an ageing clock out of proteins and then compared it with older ways of estimating biological age. Its Proteomic Aging Clock, the study reported, “showed stronger age-adjusted associations with all-cause mortality and 18 incident diseases than PhenoAge, BioAge, or leukocyte telomere length.” The last item on that list is the interesting loser. It is the only one that sounds like a physical thing wearing out.&lt;&#x2F;p&gt;
&lt;p&gt;Telomeres are repetitive stretches of DNA at the ends of chromosomes. The textbook story is that they get shorter as cells keep dividing, which makes their average length in white blood cells an appealing stand-in for how used up a body is. As an age marker it works, sort of. It points in the expected direction and lines up with other ageing measures, but it keeps getting beaten by measures that combine many signals at once. Why it underperforms is not fully settled by the research here, but the explanations sort into a few groups.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-a-telomere-is-and-why-it-looked-like-the-ideal-clock&quot;&gt;What a telomere is and why it looked like the ideal clock&lt;&#x2F;h2&gt;
&lt;p&gt;Most biological age measures are, at bottom, statistical. The Horvath clock, for instance, reads chemical tags at 353 spots on the genome and predicts chronological age with a mean absolute error of about 3.6 years, according to a summary by CD Genomics. Nobody claims those 353 spots are what ageing physically is. They are where the correlation lives.&lt;&#x2F;p&gt;
&lt;p&gt;Telomere length was the opposite pitch. It arrived with a mechanism already attached: cells divide, the ends get trimmed, and eventually the cell has trouble dividing. A marker with a story is a marker people like to measure. (I write about ageing tests for a living, which mostly means reading supplementary tables about weighted sums of CpG sites, so I understand the appeal of a ruler.) Telomere length is something like the win-loss record of ageing biology. It is easy to explain, it has been tracked forever, and it keeps getting printed next to the newer statistics that outperform it.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-telomeres-lose-to-other-ageing-measures&quot;&gt;Why telomeres lose to other ageing measures&lt;&#x2F;h2&gt;
&lt;p&gt;There are, roughly, three ways to read this. The first is that telomere length is a decent marker measured badly. The second is that it is measured well enough and simply carries less information than its competitors. The third, which is the least comfortable, is about what happens when a weak marker is pointed at an intervention and asked whether the intervention worked.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-first-reading-the-ruler-wobbles&quot;&gt;The first reading: the ruler wobbles&lt;&#x2F;h2&gt;
&lt;p&gt;In the CALERIE 2 trial of long-term calorie restriction, researchers measured telomere length two ways. One was quantitative PCR, which estimates the amount of telomere DNA in a sample by copying it and tracking how fast the copies accumulate, reported as absolute telomere length. The other was a DNA methylation algorithm. Two methods for one quantity is a sign that people do not fully trust either (or that the grant had room for both).&lt;&#x2F;p&gt;
&lt;p&gt;How noisy is the PCR version, exactly? The facts behind this article do not give a number for that, so it is not settled here. What they do show is that noise is not unique to telomeres. Elysium Health, which sells a methylation-based test, says on its science page that “a single epigenetic age sample processed repeatedly can result in biological age estimates that vary up to nine years.” A separate reliability study found that PCGrimAge was the only clock to reach good biological reliability, with an intraclass correlation above 0.75 across repeated measurements.&lt;&#x2F;p&gt;
&lt;p&gt;So the fancier clocks wobble too. The point is that wobble matters more when the real differences between people are small. If two people truly differ by a little and your instrument jitters by a lot, the instrument mostly reports the jitter. That is less a failing of anyone holding the instrument than a property of instruments.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-second-reading-one-number-against-thousands&quot;&gt;The second reading: one number against thousands&lt;&#x2F;h2&gt;
&lt;p&gt;Telomere length is a single number. Actually, that undersells how odd it is. It is the averaged length of every chromosome end across a very large number of white blood cells, each cell at its own point in its own history, compressed into one figure. Calling that “one measurement” is like calling a city’s average commute one car trip.&lt;&#x2F;p&gt;
&lt;p&gt;And it is going up against clocks built from thousands of inputs. The UK Biobank proteomic clock used 2,923 proteins. A related proteomic clock reached a correlation of 0.92 with chronological age in a Chinese biobank cohort of 3,977 people and 0.94 in a Finnish cohort of 1,990. When something like that is lined up against one averaged length, the averaged length losing is not a scandal. It is arithmetic.&lt;&#x2F;p&gt;
&lt;p&gt;To be fair, telomere length does show up where you would expect it to. The proteomic study reported that “proteomic aging was associated with telomere length as a biological age-related measure,” which means the two are picking up some of the same underlying thing. In a study of virologically suppressed people with HIV, DNA methylation-predicted telomere length was significantly shorter than in people without HIV (p &amp;lt; 0.001). Group differences are real (which is a more modest compliment than it sounds).&lt;&#x2F;p&gt;
&lt;p&gt;The trouble is that group differences and individual prediction are different jobs. A marker can reliably separate two groups of hundreds of people and still tell any one of them very little. I assume that is why telomere length keeps appearing in study after study as one column among many, rather than as the headline. I can picture the planning meeting. (Invented dialogue, not from any source.) “We’ve got the DNA in the freezer already. Can we add telomeres?” “Sure, put it next to GrimAge.”&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-third-reading-the-answer-depends-on-when-you-look&quot;&gt;The third reading: the answer depends on when you look&lt;&#x2F;h2&gt;
&lt;p&gt;This is the CALERIE 2 result in more detail, and it is the uncomfortable part. Over the first year, the intent-to-treat analysis “found no significant differences in telomere length attrition between the caloric restriction and control groups.” Over the full study, from baseline to 24 months, it again found no differences in telomere length change. But during the stretch from month 12 to month 24, which the trial called the weight maintenance period, both being in the restriction group and restricting more were associated with less absolute telomere shortening.&lt;&#x2F;p&gt;
&lt;p&gt;So: nothing in year one, something in year two, nothing across both years together. I don’t know how a reader is meant to hold all three at once. One way is that the effect needed a year to show up. Another is that a noisy marker measured at several timepoints will, presumably, produce some window that looks interesting. The facts here do not say which, and the trial does not resolve it.&lt;&#x2F;p&gt;
&lt;p&gt;It must be hard to run two years of calorie restriction on healthy adults and come away with flat telomeres, so you can see why the middle twelve months got a sentence of their own. That is not a criticism of the researchers. Reporting the sub-period is the honest thing to do. It just means the headline number for telomeres depends a lot on which calendar you pick up.&lt;&#x2F;p&gt;
&lt;p&gt;Then there is the second method. CALERIE also used DNAmTL, and so did a study of prenatal and childhood lead exposure, and so did a ten-year HIV cohort that assessed eight epigenetic clocks “plus a telomere length estimator.” In plain words, DNAmTL works like this: you skip the telomere, look at chemical tags elsewhere on the DNA that tend to go along with long or short telomeres, and make an educated guess. Its official name is DNA methylation-derived telomere length. I mean, it is a sensible tool, and it runs on the same chip as the clocks, so it costs a study nothing extra to report.&lt;&#x2F;p&gt;
&lt;p&gt;In those studies, telomere length is one more output of the methylation chip that runs the clocks that beat it, and nobody looks at a telomere at all.&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>The TRIIM Trial and the Claim of Reversing Epigenetic Age</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/triim-trial/"/>
        <id>https://bioclocktest.com/triim-trial/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/triim-trial/">&lt;p&gt;One thing that happens when someone invents a new kind of ruler is that the first person to measure something shrinking with it becomes famous, and everyone after that gets to argue about the ruler. The TRIIM trial is that first measurement for epigenetic age. It gave a small group of volunteers a year of recombinant human growth hormone (rhGH), dehydroepiandrosterone (DHEA) and metformin, read their blood with epigenetic clocks, and reported that their estimated biological ages went down.&lt;&#x2F;p&gt;
&lt;p&gt;The numbers: after 12 months, mean epigenetic age was about 1.5 years lower than at baseline, and the change in epigenetic age relative to chronological age was about minus 2.5 years compared to no treatment. On GrimAge, one of the four clocks used, the drop was about 2 years versus chronological age, and it was still there six months after treatment stopped. The paper, titled “Reversal of epigenetic aging and immunosenescent trends in humans,” called itself “the first report of an increase in predicted human lifespan via a currently accessible aging intervention, based on epigenetic age estimation.” It enrolled nine volunteers.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-the-triim-trial-measured-and-when&quot;&gt;What the TRIIM trial measured, and when&lt;&#x2F;h2&gt;
&lt;p&gt;The name stands for Thymus Regeneration, Immunorestoration, and Insulin Mitigation, which suggests the clocks were sort of a passenger and the immune system was driving. (The immune results are the other half of the paper’s title, and I am going to leave them to the paper.) The drug list is notable mostly for growth hormone, a substance with a long and colorful reputation in certain locker rooms, all of it presumably about thymus research.&lt;&#x2F;p&gt;
&lt;p&gt;The timing is the interesting part. From months 0 to 9, the reported rate of epigenetic aging was minus 1.6 years per year. From months 9 to 12 it accelerated to minus 6.5 years per year. So most of the action came late. Three months is three months.&lt;&#x2F;p&gt;
&lt;p&gt;Here is a hypothetical, invented and not a participant. A 60-year-old reverses at minus 6.5 years per year and never stops. Around age 69 his blood reads as a newborn’s. Around 70 it reads as someone who has not been born yet, which is a Benjamin Button plot the studio would have sent back as too much. Nobody claimed this, to be fair. But “predicted lifespan” is made of extrapolation (that is what the word predicted is doing), and the slope is the slope.&lt;&#x2F;p&gt;
&lt;p&gt;One more item belongs in the record, and it should be entered formally. Be it noted that at the commencement of the study, before any person received any drug, the participants’ epigenetic ages were, on average, lower than their chronological ages. That is to say: the instrument had already declared the volunteers younger than their birth certificates, prior to and independent of the intervention whose effect it was about to measure.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;the-obvious-objection-it-was-nine-people&quot;&gt;The obvious objection: it was nine people&lt;&#x2F;h2&gt;
&lt;p&gt;The sensible reaction is that nine people is not a trial, it is a dinner party, or a book club or whatever. That reaction is correct. With nine people, one odd blood draw moves the average a lot. And with three drugs given together, there is no way to say which one did anything. The original study did not have control groups for either question (the follow-up’s registry listing is quite direct about this, in the polite way registries are direct).&lt;&#x2F;p&gt;
&lt;p&gt;To be fair, small first studies are how a lot of medicine starts: someone tries a thing on a handful of people, writes down what happened, and asks for money to try it on more. The problem is not that TRIIM was small. It is that its headline was not small.&lt;&#x2F;p&gt;
&lt;p&gt;But the point is that a small sample is the boring objection, because it is fixable. Run it with 900 people. The stranger problem survives that.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-the-comparison-is-stranger-than-the-headcount&quot;&gt;Why the comparison is stranger than the headcount&lt;&#x2F;h2&gt;
&lt;p&gt;Look at the two headline numbers again. Mean epigenetic age fell about 1.5 years from baseline. The effect “compared to no treatment” was about 2.5 years. The gap between them is one year, which is also how long the study lasted. The arithmetic suggests that “no treatment” here means, in effect, the calendar: an untreated person is assumed to get one epigenetic year older per calendar year, and anything less counts as the drugs working. (I have checked my own calendar, and it is very consistent.)&lt;&#x2F;p&gt;
&lt;p&gt;Is that a crazy assumption? Not exactly. It is what the clocks were built to do on average. But it means the control arm, in a study without a control arm, is the clock’s promise to behave, and the thing being measured is whether the clock kept its promise. If the clock wanders on its own, the wandering shows up as effect.&lt;&#x2F;p&gt;
&lt;p&gt;I mean, picture the lab meeting. “Hey, good news, everyone got younger.” “Did we run them twice?” “Why would we run them twice?”&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-much-a-clock-wobbles-on-its-own&quot;&gt;How much a clock wobbles on its own&lt;&#x2F;h2&gt;
&lt;p&gt;Quite a lot, depending on whose numbers you use. For instance, one test company reports that technical replicates of published clocks (Horvath, Hannum, PhenoAge) produced deviations of three to nine years between measurements of the same sample. That is one company’s process, not the trial’s lab, and averaging across nine people narrows noise somewhat, so I don’t know how much of it applies here. The facts don’t settle it. But the reported effect was 2.5 years. The reported wobble goes to nine.&lt;&#x2F;p&gt;
&lt;p&gt;There is a second way to see the scale. The Horvath pan-tissue clock, which uses 353 CpG sites, has a mean absolute error of about 3.6 years when predicting chronological age across tissues, according to CD Genomics. So the clock’s ordinary miss on any one person is bigger than the trial’s average effect. That is not a contradiction (an average over nine people can be more precise than any single reading), but it gives a sense of how finely the instrument is being asked to cut. The Hannum clock, another one in the follow-up, uses 71 CpG sites and is built for blood.&lt;&#x2F;p&gt;
&lt;p&gt;There is also no shared yardstick. The Scientist reported that as of 2026 the field “has not yet achieved reference standards,” meaning you cannot send one sample to several validated labs and get the same age back. A 2026 systematic review of geroscience intervention trials found no comparative study where clock changes and changes in patient-important outcomes lined up uniformly. (The CALERIE calorie restriction trial, covered elsewhere on this site, found clock effects that were small.)&lt;&#x2F;p&gt;
&lt;p&gt;You and I both know what happens to a number like “minus 6.5 years per year” once it leaves a paper. It does not stay in the paper. I say this as someone whose job is mostly reading sentences containing the word “acceleration” and who still could not tell you his own GrimAge.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-triim-x-is-supposed-to-fix&quot;&gt;What TRIIM-X is supposed to fix&lt;&#x2F;h2&gt;
&lt;p&gt;The follow-up is TRIIM-X, registered as NCT04375657. Intervene Immune’s program page says it uses multiple clocks including Horvath, Hannum, PhenoAge and GrimAge, and the registry listing says it includes control groups to test biomarker variability and the contribution of individual medications. That is, more or less exactly the two holes. A control group, for the record, is the part of a study that gets older on purpose.&lt;&#x2F;p&gt;
&lt;p&gt;Its stated objective is “to obtain information for designing an effective personalized and adaptive treatment regimen for a larger and more diverse study population.” Translated flatly: this study exists to help design the next study, which will have more people and let the recipe vary by person. It is an honest sentence, and also the most clinical-research sentence ever written, lol.&lt;&#x2F;p&gt;
&lt;p&gt;The facts here include no results from it yet. Whether the reversal holds up once there are controls is not settled. The measurement, meanwhile, comes with its own caveat in writing.&lt;&#x2F;p&gt;
&lt;p&gt;Elysium Health’s science page notes that a single epigenetic age sample processed repeatedly “can result in biological age estimates that vary up to nine years.”&lt;&#x2F;p&gt;
</content>
        
    </entry>
    <entry xml:lang="en">
        <title>TruAge and TruDiagnostic: What the Test Reports</title>
        <published>2026-11-01T00:00:00+00:00</published>
        <updated>2026-11-01T00:00:00+00:00</updated>
        
        <author>
          <name>
            
              Unknown
            
          </name>
        </author>
        
        <link rel="alternate" type="text/html" href="https://bioclocktest.com/truage/"/>
        <id>https://bioclocktest.com/truage/</id>
        
        <content type="html" xml:base="https://bioclocktest.com/truage/">&lt;p&gt;TruDiagnostic says its test reads more than 1 million sites on your DNA. Then it hands back 64 numbers. That is a pretty aggressive compression ratio, and most of what is interesting about the product lives in the space between the million and the 64.&lt;&#x2F;p&gt;
&lt;p&gt;When people type “truage” into a search box, they mean TruDiagnostic’s epigenetic age test. Here is the plain answer. The product combines three epigenetic clocks, called OMICmAge, SymphonyAge and DunedinPACE, and the company says the report covers 64 “longevity biomarkers” across biological age, aging speed and organ-specific aging for 11 body systems, including brain, heart and liver. DunedinPACE has a published academic record you can read on its own terms. For the other two, the company reports “50+ peer-reviewed publications” supporting its products and more than 80 academic and research partnerships, but those counts do not by themselves settle how well each clock holds up. On that question the record is not settled.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-the-truage-report-contains&quot;&gt;What the TruAge report contains&lt;&#x2F;h2&gt;
&lt;p&gt;Here is what an epigenetic clock does, stated as plainly as I can. It looks at a lot of spots on your DNA, notes which ones carry a particular chemical tag, compares the pattern to a large group of people whose ages and health are known, and guesses something about you. That is called an epigenetic clock. (The chemistry has its own article on this site, and so does the general question of what these tests measure.)&lt;&#x2F;p&gt;
&lt;p&gt;TruAge runs three of these. OMICmAge, in TruDiagnostic’s description, is a “multi-omic-informed methylation clock” that compares your biological age to your chronological age. SymphonyAge estimates a biological age for each of 11 organ systems. DunedinPACE does not give you an age at all. It gives you a speed: how fast you are aging, as a rate.&lt;&#x2F;p&gt;
&lt;p&gt;So the report is less like a single verdict and more like the back of a baseball card, with an overall number, a batting average and a column of splits by organ. That’s fine. A baseball card is a useful object. It does help to know which stats have been around long enough to mean something, though.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;why-one-test-needs-three-clocks&quot;&gt;Why one test needs three clocks&lt;&#x2F;h2&gt;
&lt;p&gt;The chain starts with researchers, who get rewarded (in grants, citations and tenure, the usual currency) for building a new clock that does something the old ones did not. The Horvath clock, published in 2013, was the landmark: it predicted age across many tissues and cell types. The Hannum clock was built on 482 people aged 19 to 101, all of Caucasian or Hispanic ethnicity. Both picked their DNA sites using a statistical method called elastic net regression, aimed at producing the most accurate estimate of chronological age.&lt;&#x2F;p&gt;
&lt;p&gt;This creates a weird problem. A clock that perfectly predicts your chronological age is perfectly useless to you. You already know your birthday. The interesting part is the error, the gap between what the clock says and what your driver’s license says, and the next generation of clocks was trained to make that gap mean something (risk of disease, rate of decline, etc.).&lt;&#x2F;p&gt;
&lt;p&gt;The accuracy race also keeps going on its own track. One 2022 study built clocks with new feature-selection methods that beat the Hannum, Horvath and Weidner clocks on two outside datasets, and its most predictive clock used only 35 DNA sites, with an R-squared of 0.87. Thirty-five. TruDiagnostic reads more than a million. Those are different jobs, I suppose, but it is worth knowing that a long list of sites is not the same thing as a better answer.&lt;&#x2F;p&gt;
&lt;p&gt;Now the company link. A company selling a test wants its report to be fuller than the next company’s, and organ ages are a natural way to get there. (“Hey, what if the brain had its own number,” is a sentence I am inventing, but somebody in product said a version of it somewhere.) Elysium Health’s Index test, for comparison, reports ages for nine body systems. TruAge reports 11. More rows on the card.&lt;&#x2F;p&gt;
&lt;p&gt;In fairness, TruDiagnostic at least names its clocks, which is more than the field usually manages. As The Scientist put it, “most companies offering biological age test kits do not detail which epigenetic clock they use.”&lt;&#x2F;p&gt;
&lt;h2 id=&quot;how-dunedinpace-was-built-to-repeat-itself&quot;&gt;How DunedinPACE was built to repeat itself&lt;&#x2F;h2&gt;
&lt;p&gt;Of the three clocks, DunedinPACE is the one designed most directly around a measurement problem. Its predecessor, DunedinPoAm, was published by Belsky and colleagues in 2020 as a single-time-point blood test of the pace of aging. DunedinPACE kept the idea and changed the inputs: its builders threw out DNA sites with low test-retest reliability before training, then ran elastic net on what was left. The result showed reliability above 0.90.&lt;&#x2F;p&gt;
&lt;p&gt;Why does that matter? Because a clock you take twice should say roughly the same thing twice. Elysium Health (which sells a competing test, so read its numbers with that in mind) reported that technical replicates of the Horvath, Hannum and PhenoAge clocks deviated by three to nine years between measurements of the same sample, and that its own Index test agreed within zero to 1.5 years. A separate platform called TranslAGE evaluated the technical and biological reliability of 18 epigenetic clocks. The reliability question has its own article on this site; the short version is that some clocks wobble more than others, and DunedinPACE was engineered specifically not to.&lt;&#x2F;p&gt;
&lt;p&gt;There is also a sample detail. Almost all clocks in commercial tests are trained on blood. A test that matches its sample type to its training data is, at least, not fighting its own inputs.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;what-published-validation-shows-for-truage-and-its-clocks&quot;&gt;What published validation shows for TruAge and its clocks&lt;&#x2F;h2&gt;
&lt;p&gt;Here is where the chain gets interesting. The AMA Journal of Ethics, in a December 2025 commentary, says these tests have been shown to be fairly reliable indicators of a person’s risk of developing certain age-related diseases. That is a real finding about association. People with faster clocks, on average, get sicker.&lt;&#x2F;p&gt;
&lt;p&gt;On the other hand, the thing a customer most wants to know is different. The customer wants to know whether making the number go down makes them healthier. A 2026 systematic review of geroscience intervention trials looked for exactly that, trial-level validation that clocks work as surrogate endpoints, and found none.&lt;&#x2F;p&gt;
&lt;p&gt;This is not a small distinction. The TRIIM trial used four epigenetic clocks and claimed to be the first report of an increase in predicted human lifespan from a currently accessible intervention, based on epigenetic age estimation. Its follow-up, TRIIM-X, added control groups to test biomarker variability and the contribution of individual medications, which the original did not have. So the field is aware of the gap. It is just still standing in it.&lt;&#x2F;p&gt;
&lt;p&gt;And the incentive at this link is clear. Anyone selling an intervention wants a number that moves. A clock that moves is a clock you can sell twice.&lt;&#x2F;p&gt;
&lt;h2 id=&quot;where-the-incentive-chain-ends-up&quot;&gt;Where the incentive chain ends up&lt;&#x2F;h2&gt;
&lt;p&gt;Everyone in this chain is acting sensibly from where they stand. Researchers publish better clocks because that is the job. Companies bundle several clocks and add organ ages because a fuller report is a better product. Customers buy the test because a number about your own aging is interesting, and doctors look at the results because patients bring them in.&lt;&#x2F;p&gt;
&lt;p&gt;Anyway, the AMA commentary describes where this can arrive. A healthy 56-year-old, identified as AA, received a direct-to-consumer result (not identified as a TruAge result) saying he was “biologically 65-years-old.” He then followed the manufacturer’s recommendations and began courses of the company’s brain health and cellular repair supplements. The test and the supplements were, in the corporate-structure sense, related.&lt;&#x2F;p&gt;
&lt;p&gt;The commentary’s view is that these tests cannot tell us how old a person “really” is, and it calls biological age a fiction. It asks companies to be clearer about how the tests work, how reliable they are and what they can and cannot show, and warns that results can cause psychological harm and make age-related discrimination look justified. That is a strong position. I guess you can read it as the last link of the chain looking back at the first.&lt;&#x2F;p&gt;
&lt;p&gt;Nobody in the chain set out to produce that patient. Each link just got paid for its own piece: the clock, the report, the number, the next step.&lt;&#x2F;p&gt;
&lt;p&gt;A million sites go in. Sixty-four numbers come out. The compression was never the aggressive part.&lt;&#x2F;p&gt;
</content>
        
    </entry>
</feed>
