Deep divefrom Episode 3Longevity1 September 202616 min

Responsiveness of epigenetic aging biomarkers to longevity i

A deep dive into one paper: what was known before, what the researchers did, what the numbers say and what they don't.

The paper: Responsiveness of epigenetic aging biomarkers to longevity interventions in humans (Nature Medicine)

Chapters

Transcript

The script the two AI voices read from. The times are worked out from the script, not heard, so a line can start a few seconds early or late.

Read the whole episode

Cold open

Host 1Welcome to the Hastrup Review of Science — where artificial intelligence reads real science, so you don't have to pretend you understood the paper. One paper today. A group has assembled the largest database of longevity trials ever harmonized, and asked a deceptively simple question: do our aging clocks actually respond when we try to slow aging down?

Host 2Sehgal, Smith and colleagues have curated TranslAGE, a harmonized database of fifty one longitudinal interventional studies, calculated sixteen prominent epigenetic clocks consistently across all of them, and discovered which clocks respond to which interventions and why. The paper is published in Nature Medicine.

What is an epigenetic clock

Host 1The paper is called Responsiveness of epigenetic aging biomarkers to longevity interventions in humans. Before we get into what they found, we need to set the stage. What is an epigenetic clock?

Host 2As cells age, the chemical tags on their DNA change in predictable ways. Specifically, methyl groups are added or removed at cytosine bases in CpG dinucleotides. In 2013, Steve Horvath demonstrated that you could train a machine learning model on the methylation levels at a few hundred of these sites, and it would predict a person's chronological age with remarkable accuracy — within about three and a half years.

Host 1So the clock reads your DNA methylation pattern and tells you how old you are biologically.

Host 2That was the original idea. The Horvath clock was the first pan-tissue epigenetic clock. Around the same time, Hannum and colleagues built a similar clock optimized for blood samples. These are often called first-generation clocks. They were trained to predict chronological age, which means they are very good at telling you how many birthdays you have had, but not necessarily how fast you are aging.

Host 1There is a difference between knowing your age and knowing your rate of aging?

Host 2A critical difference. Two sixty-year-olds can have very different biological trajectories. One might have the cardiovascular system of a fifty-year-old, the other of a seventy-year-old. First-generation clocks struggle to capture that distinction because they were optimized for the wrong target. They learned to count birthdays, not to measure physiological decline.

Second and third generation clocks

Host 1So what came next?

Host 2Second-generation clocks. In 2018, Levine and colleagues developed PhenoAge, trained not on chronological age but on a composite of clinical biomarkers — albumin, creatinine, glucose, C-reactive protein, lymphocyte percentage and several others — that predict mortality risk. In 2019, Lu and colleagues built GrimAge, trained directly on mortality outcomes. These clocks use DNA methylation to estimate surrogate proteins and clinical markers that are associated with death.

Host 1And then a third generation?

Host 2In 2022, Belsky and colleagues at Duke developed DunedinPACE — the Pace of Aging Computed from the Epigenome. Instead of estimating your biological age at a single time point, DunedinPACE was trained on longitudinal data from the Dunedin birth cohort in New Zealand. It measures the rate of biological aging in years of physiological decline per calendar year. A score of one means you are aging at the expected rate. A score of 1.2 means you are aging twenty percent faster.

Host 1So you have clocks that predict age, clocks that predict mortality, and clocks that predict the pace of aging. Why does it matter which one you use in a clinical trial?

Host 2Because the longevity field has a fundamental problem. If you want to prove that an intervention extends human lifespan, you need a trial that runs for decades. That is prohibitively expensive, logistically impractical, and ethically complex. The hope has been that epigenetic clocks could serve as surrogate endpoints — measure someone's biological age before and after an intervention, and if the clock moves in the right direction, declare the intervention effective.

The problem

Host 1But nobody had systematically tested whether the clocks actually respond to interventions.

Host 2Exactly. Individual studies have reported that a specific clock moved after a specific intervention. The CALERIE caloric restriction trial showed changes in DunedinPACE. The TRIIM trial of growth hormone plus DHEA plus metformin reported GrimAge reversal. But each study used different clocks, different processing pipelines, and different statistical methods. There was no way to compare results across studies or to ask the broader question: which clocks respond to which types of interventions?

TranslAGE

Host 1And that is the gap this paper fills. Tell me about TranslAGE.

Host 2TranslAGE is a harmonized database that took fifty one longitudinal interventional studies — both public datasets and private ones contributed by collaborating researchers — and ran all of them through the same processing pipeline. They collected raw DNA methylation array data, applied the same quality control, the same normalization, and then calculated sixteen epigenetic clocks plus ninety four additional DNA methylation biomarkers for every sample.

Host 1Sixteen clocks across fifty one studies. What kinds of interventions are represented?

Host 2A wide range. Pharmacological interventions such as rapamycin, metformin, growth hormone, vitamin D, and nicotinamide mononucleotide. Lifestyle interventions including caloric restriction, exercise programs, meditation, and dietary changes. Surgical interventions like bariatric surgery. And combination programs that mix several approaches. The study sizes range from small pilot trials to large randomized controlled studies.

Key findings

Host 1So you have a landscape view for the first time. What did they find?

Host 2The first major finding is that not all clocks are equally responsive. First-generation clocks — the ones trained on chronological age, like the original Horvath clock — showed almost no consistent response to any intervention. They barely moved.

Host 1The most famous clock in the field is essentially unresponsive.

Host 2To interventions, yes. Which makes mechanistic sense. First-generation clocks were trained to predict birthdays. The methylation sites they rely on track developmental and maintenance programs that tick along steadily regardless of whether you are taking rapamycin or running marathons. They are measuring something real about age, but it is something that interventions do not easily alter.

Host 1And the later-generation clocks?

Host 2Clocks trained on mortality or pace of aging showed the strongest and most consistent responses. GrimAge and DunedinPACE were the standout performers. They responded across multiple intervention types, and when both were measured in the same study, they tended to agree in direction and rough magnitude. That concordance across independently developed clocks is encouraging because it suggests they are capturing a real biological signal rather than idiosyncratic noise.

Why some clocks respond

Host 1What drove the difference? Is it just a matter of what they were trained on?

Host 2Partly, yes. Training target matters enormously. But there is a deeper point. GrimAge is constructed as a composite of DNA methylation surrogates for plasma proteins — things like plasminogen activator inhibitor one, growth differentiation factor fifteen, cystatin C, and smoking pack-years estimated from methylation. Each of those surrogates reflects a modifiable physiological process. When an intervention changes inflammation, metabolic function, or cardiovascular risk, the methylation surrogates for those proteins shift, and GrimAge moves.

Host 1So GrimAge is responsive because its components are biologically connected to the things interventions actually change.

Host 2Precisely. And this leads to the paper's second major finding: clocks with multiple subscores — what they call explainable clocks — provide substantially more insight than single-score clocks. When GrimAge moves, you can decompose the change and see which subscore drove it. One intervention might primarily reduce the inflammation component. Another might shift the metabolic signature. A third might affect the smoking-related methylation surrogate even in non-smokers, which points to shared pathway activation.

Host 1That is a meaningful advance over just getting a single number that goes up or down.

Host 2It is. A single-score clock tells you the intervention did something. An explainable clock tells you what it did, to which biological systems, and that mechanistic specificity is what you need to design better interventions and to understand why some work and others do not.

A word from our sponsor

Host 1Does your scalp seem less populated than it used to be? Do you find yourself describing your forehead as larger rather than balder?

Host 2The Hastrup Review of Science is brought to you by hair-loss shampoo. Caffeine-infused. Tingly. Optimistic.

Host 1I want to be clear: we are not claiming this shampoo reverses epigenetic aging of the scalp. We are saying it makes the shower feel purposeful.

Host 2I gave some to my father. He said it did not bring his hair back, but it did bring back his confidence. Those may be the same thing.

Host 1Hair-loss shampoo. Because your follicles deserve an intervention, even if the clock keeps ticking. Our referral link is in the show notes.

Which interventions work

Host 1Back to the findings. They said pharmacological and lifestyle interventions drove the strongest responses. Can you break that down?

Host 2Among the intervention categories they examined, pharmacological and lifestyle approaches consistently produced larger and more statistically significant changes in the responsive clocks compared to supplements or single-nutrient interventions. Caloric restriction and multi-component lifestyle programs tended to produce broad effects across multiple clock subscores, whereas some pharmacological interventions produced sharper, more targeted changes in specific components.

Host 1Did any intervention stand out as a clear winner?

Host 2The paper is deliberately careful about ranking individual interventions because the studies vary enormously in sample size, duration, population characteristics, and dosing. What they emphasize instead is the pattern: interventions that engage multiple physiological systems simultaneously tend to produce broader clock responses than those targeting a single pathway. That is consistent with aging being a multi-system process rather than a single-target disease.

Study design matters

Host 1They also found that study design matters. What role do population characteristics play?

Host 2This is one of the most practically important findings. The characteristics of the study population — age, baseline health, sex — and the study duration were key factors in determining whether an intervention produced a detectable clock response. Older participants with more room for biological improvement tended to show larger responses. Shorter studies often failed to detect changes that longer studies captured.

Host 1So if you run a six-week trial in healthy thirty-year-olds, you are unlikely to see the clock move.

Host 2Correct. And that has implications for how the field has been interpreting negative results. A study that found no epigenetic clock change after a short intervention in young healthy volunteers may not mean the intervention does not work. It may mean the study was not designed to detect the signal. This paper provides guidance on the minimum study duration and population characteristics needed to have statistical power to detect real changes.

Caveats

Host 1That is genuinely useful for trial design. What are the major caveats?

Host 2Several important ones. First, clock responsiveness is not the same as clinical validity. Just because a clock moves in response to an intervention does not prove the intervention is actually slowing aging. The clock could be responding to a transient physiological change that has nothing to do with longevity. To validate these clocks as true surrogate endpoints, you ultimately need to show that clock changes predict actual health outcomes — reduced disease incidence, extended healthspan, lower mortality.

Host 1Which requires the long follow-up studies that the clocks were supposed to replace.

Host 2There is a circularity, yes. Clocks are proposed as shortcuts to avoid long trials, but proving they are valid shortcuts requires long trials. This is the fundamental surrogate endpoint validation problem in all of medicine, not unique to aging research. The FDA requires rigorous evidence that a surrogate actually predicts the outcome it is substituting for.

Host 1Second caveat?

Host 2The fifty one studies in TranslAGE vary enormously in quality. Some are well-powered randomized controlled trials. Others are small, uncontrolled pilot studies. The database harmonizes the methylation processing but cannot harmonize the underlying study design. Results from a twelve-person open-label trial are weighted alongside results from a two-hundred-person RCT, and that heterogeneity makes aggregate patterns harder to interpret.

Host 1Third?

Host 2All of the methylation data comes from blood samples, specifically from peripheral blood leukocytes. Blood is convenient to collect, but aging is a tissue-specific process. Your brain, your liver, and your muscle tissue may age at different rates and respond to interventions differently. A clock measured in blood captures systemic signals, particularly immune and inflammatory ones, but it may miss organ-specific effects entirely.

Host 1Fourth?

Host 2The ninety four additional DNAm biomarkers used to decompose clock responses are themselves derived from statistical associations. When the paper says an intervention shifted the inflammation subscore of GrimAge, that subscore is a methylation surrogate for a plasma protein, not a direct measurement of the protein itself. There is an inference chain — from intervention to methylation change to estimated protein level to biological interpretation — and each link introduces noise and potential misattribution.

What this means

Host 1Where does this leave the field? What is the practical impact?

Host 2Three concrete contributions. First, TranslAGE provides a standardized reference that any future trial can benchmark against. If you run a new rapamycin study, you can compare your clock results to every other study in the database using the same processing pipeline. That alone eliminates a major source of confusion in the current literature.

Host 1Second?

Host 2It provides evidence-based guidance for clock selection. If you are designing a trial and need to choose which clocks to measure, this paper tells you to prioritize mortality-trained and pace-of-aging clocks, and ideally to use explainable clocks with decomposable subscores. First-generation chronological age clocks should not be used as primary endpoints in intervention studies.

Host 1And third?

Host 2It lays the groundwork for reducing the multiple-testing burden. There are dozens of epigenetic clocks and hundreds of DNAm biomarkers. Testing all of them in every study creates a severe statistical correction problem. This paper identifies which subset is most likely to respond, allowing future trials to pre-specify a smaller panel and gain statistical power without inflating false positive rates.

Bottom line

Host 1Bottom line: what should someone following longevity research take away from this?

Host 2That the epigenetic clock field is maturing from a period of proliferation — where new clocks were published faster than anyone could compare them — into a period of systematic validation. This paper does not tell us whether any specific intervention slows aging. What it tells us is which measurement tools are sensitive enough to detect a signal if one exists, and how to design studies that give those tools the best chance of working. It is infrastructure, not a breakthrough. But it is the kind of infrastructure the field urgently needed.

Sign-off

Host 1That is the deep dive. Responsiveness of epigenetic aging biomarkers to longevity interventions in humans, from Sehgal, Smith and colleagues, published in Nature Medicine. Link in the show notes.

Host 2This has been the Hastrup Review of Science. See you next week.