I think I know why methylation clocks are failing

Methylation clocks promise to revolutionize testing of anti-aging interventions in humans, by measuring their effectiveness in a few months instead of having to wait years to see if they affect mortality statistics. But there are signs that the clocks we have can be deceptive. I think this is because the clocks capture aspects of aging that are defensive and aspects that are programmed, and there has been no effort to tease them apart. To the extent that the body is defending itself against perceived damage, the clocks are measuring the wrong thing. (This summary probably won’t make sense until you read the details below — please be patient.


I recently submitted an academic study on this subject.

We have a lot of life extension treatments that work in lab animals. Do any of these work in humans? In practice, we need aging clocks to test them, because testing them directly requires tens of thousands of people to be followed over decades for each intervention. I’ve been enthusiastic about aging clocks since Steve Horvath published his groundbreaking analysis in 2013. But recently I’ve realized that the field needs a course correction.

Epigenetic clocks are based on patterns of methylation in our DNA. Genes are turned on and off at different places in the body, at different times of day, and, crucially, at different stages of life. Methylation is the most accessible and easiest to measure of many methods by which the body turns genes on and off. By focusing on the methylation patterns that change consistently over a lifetime, Hannum, Horvath and others who came after them have created computer algorithms that can calculate a “biological age”.  

The idea of “biological age” doesn’t depend on your theory of aging. But the utility of epigenetic clocks in assessing the benefits of putative anti-aging measures certainly does depend on fundamental concepts about aging, about which experts are still divided.

Why does gene expression change in old age?

“Nothing in biology makes sense except in the light of evolution.”
— T. Dobzhansky

1)    If you believe in programmed aging, then the directed changes in gene expression are means of self-destruction. Genes are turned on that increase inflammation, destroying arteries and neurons. Apoptosis is up-regulated to the point where healthy muscle and brain cells are dying. Protective anti-oxidants, DNA repair, and autophagy are down-regulated. All this destruction is accomplished via turning genes on and off. If any intervention sets back the methylation clock, then there is less self-destruction, more repair and maintenance. We expect that the body will live longer if the methylation clock reads a lower age.

2)    If you believe the neo-Darwinist theory that the body cannot be purposely destroying itself, then aging is an accumulation of incidental damage at the cellular and molecular levels. If there are associated epigenetic changes, these cannot be causing the destruction, so they must be a response to the damage. Changes in gene expression as captured in the methylation clocks must be the body’s effort to protect itself with increased immune function, increased autophagy, increased antioxidants, increased DNA repair. If any intervention sets back the methylation clock, then there is less repair and maintenance. We expect that setting the aging clock back to a younger age will actually decrease life expectancy. This insight is counter-intuitive, but, if correct, it changes the logic of methylation clocks. 

For people who don’t believe that aging is an evolved program, the whole idea of a methylation clock is a non-starter. No matter how accurate the clock is, setting it back is counter-productive. Even if the clock is calibrated to markers of health (like the PhenoAge clock) or calibrated to actual mortality (GrimAge), it is still based on the body’s response to damage, and not on the damage itself. Setting back the clock is counter-productive, because it means dialing down the body’s repair and maintenance system. 

Since 2013, there has been a kind of double-think in the world of anti-aging research. Most researchers, at least in public, continue to embrace perspective (2), even as they adopt methylation clocks to evaluate the interventions they develop.

All this is assuming perspective (2). But I’m notorious for being a proponent of perspective (1). From perspective (1), turning back the methylation clock is a good thing. It means that the body’s program of self-destruction is dialed back. So where’s the problem?

In recent years, I have become convinced that epigenetic changes of both types (1) and (2) are taking place simultaneously as the body ages. The body is at war with itself. The self-destructive adaptations listed above are real: dialing down repair and maintenance, promoting systemic inflammation, apoptosis of healthy cells, derangement of the immune system. But the body retains its protective responses, and there are also changes in gene expression that ramp up the repair processes. All the present clocks include a mixture of (1) and (2); this is why we do not yet have a reliable metric for the efficacy of anti-aging technologies. 

What is the evidence that changes of types (1) and (2) are both components of all extant aging clocks?

Some of the best-established interventions for extending lifespan do not affect the major algorithmic clocks, or do so modestly compared to what might be expected from their observed effects on lifespan. Rapamycin extends lifespan of male mice without affecting their methylation age in the Horvath rodent clock. Participants in the CALERIE study who have adopted extreme CR diets showed no significant benefit according to either the GrimAge or PhenoAge clocks.

Conversely, Katcher’s intravenous infusion of exosomes (E5) has a dramatic effect on the Horvath rodent/human clock, reducing epigenetic age by half, but thus far seems to extend lifespan less than the clock setback would imply. The Conboys recently published a withering criticism of the utility of current methylation clocks, and of the machine learning algorithms from which they are created. They report that clocks in common use do not respond as expected to known life-shortening conditions, such as Down Syndrome, inflammaging associated with arthritis, and Parkinson Disease.

Here’s what first clued me in

The GrimAge clock of Lu and Horvath was trained on actual mortality data, using historic blood samples for which the future history of the donors was known. This was a major advance from previous clocks. But one element of the GrimAge development alerted me to the issue concerning type (2) changes, as described above.

Part of the training of GrimAge involved a methylation image of the subject’s smoking history. Smoking is known to accelerate aging and shorten life expectancy. Certain patterns of methylation are associated with smoking, and are also valuable predictors of time until death. These were included in the GrimAge algorithm.

My assumption was that smoking decreases longevity by damaging tissue of the lungs, not by turning on the program of self-destruction. Therefore, if there are methylation changes associated with smoking, they are probably of type (2). In other words, the methylation signature of a smoker who scores as “older” in GrimAge is likely to include activation of more protective pathways than a non-smoker who scores “younger”.

This is an important clue. The methylation profile of a smoker is useful in constructing a GrimAge clock, but it should be counted in reverse. Methylation changes associated with smoking are statistically associated with shorter lifespan, but mechanistically with protection. These changes should have been included in algorithmic clocks with negative coefficients, signaling a younger biological age. This was not how the GrimAge clock was constructed in fact. Methylation changes associated with smoking were included in the GrimAge clock with positive coefficients (simply because they are statistically associated with a shorter life expectancy).

In general, the methylation image of smokers is an example of type (2). All type (2) changes should be counted with negative coefficients in methylation clocks, even though they are statistically associated with older ages and shorter remaining life expectancy.

So it is crucial to distinguish epigenetic changes of type (1) from type (2)

 The story of GrimAge carries a message that suggests ways that methylation changes of types (1) and (2) might be teased apart in algorithmic clocks. Present clocks don’t distinguish between (1) and (2) so presumably the two types of methylation changes are combined in a way we might connote as (1) + (2). The goal would be to create a clock built on type (1) changes alone, or, more speculatively, penalize the clock for type (2) changes, so with the result that the algorithm measures (1) – (2) rather than (1) + (2).

The long-term goal would be to understand the metabolic consequences of each CpG change, separately and in combination, so that a clock could be constructed with full confidence that it scores beneficial and detrimental methylation changes appropriately.

Lacking this understanding in the interim, we might make progress toward distinguishing (1) and (2), by learning from the smoking example. One way to acquire a database of type (2) changes is that animal models might be injected with pro-inflammatory cytokines, and their epigenetic consequences mapped. Since the inflammation is imposed externally, we should presume that the response is all type (2). 

Similarly, the animals’ immune systems might be challenged, or they might be subjected to laceration or small doses of radiation, again to chart the epigenetic response to compile a list of candidates for type (2) changes. These experiments could not ethically be performed on humans, however there are humans whose aging is accelerated by non-epigenetic factors, including alcohol and drug abuse. Such people might be tested as part of the quest for type (2) changes. People healing from physical and emotional trauma might also be presumed to have epigenomes modified in the direction of type (2).

Hormesis is the body’s overcompensation to challenges. The body is damaged by something we do or we eat or suffer from, but the body overcompensates to the damage such that we live longer. 

Caloric restriction (CR) is the best-established example of hormesis. We might have most confidence in the epigenomes of people and animals subjected to caloric restriction]. Across the animal kingdom, CR is the most robust anti-aging strategy known at present, and we can be confident in subtracting CR-associated epigenetic changes from any algorithmic measure of biological age. These changes can be observed in humans, as in the CALERIE study mentioned above; or they can be observed in rodents, which have enough commonality to human metabolism that some of the same methylation sites have common functions in both. Our methylation clocks should be calibrated to be sure that the changes associated with CR are scored toward a younger biological age. 

In addition to CR, there are dozens of interventions known experimentally to extend lifespan in rodents, including juvenile exosomes, rapamycin, certain peptides, vitamin D, NAC, certain anti-inflammatories and angiotensin inhibitors. Recently, some of these have been tested for their effect on algorithmic clocks; and this has been interpreted as evidence for or against the intervention. We might reverse the logic and interpret the same data as training to calibrate clocks, assuming that these changes must be beneficial, and the clock algorithm should reward them with a younger age. If an intervention is known to increase lifespan, then we may presume that epigenetic changes observed in response to that intervention are beneficial.

Before 2013, biological age was estimated with measures of performance and appearance: grip strength, gait speed, athletic endurance, memory, exhalation volume (FEV), skin wrinkles, arterial inflammation, cartilage integrity. In the age of epigenetics, these physical characteristics retain their value as predictors of mortality, and a hybrid clock might be devised, combining physical and epigenetic factors. 

For the future

I know of projects at Stanford (the Biomarkers of Aging Consortium), at Tally Health, and at Tru Diagnostic to develop the next generation of clocks to evaluate anti-aging interventions. There are probably other, parallel efforts that I don’t know about, but I will hear about them in your comments. I hope that the theory and suggestions in this blog may be useful to them.

37 thoughts on “I think I know why methylation clocks are failing

  1. What if the body’s main clock uses damage for clues? Then, alcohol consumption or smoking could actually still be acting through the epigenetic program.

    That would explain why Sinclair’s rats with induced double-stranded DNA breaks (DSBs) seem to age faster. At the same time, as I understand it, the direct effects of DSBs aren’t very significant, so they can’t be aging rats through direct damage.

    • >>What if the body’s main clock uses damage for clues?

      This doesn’t strike me as how biology usually works. You’re saying the body detects damage and then decides to damage itself more. I’d argue that this is the opposite of hormesis. Unlikely.

  2. I think you make an assumption too far Josh that interventions that extend life are making those treated younger. They may merely be increasing robustness in some way; they may even be making them older by some measures, even though they live longer. This view more intuitively agrees with what you’d expect to see in a calorie restricted animal (or human). In short, becoming young is not the only way to live longer!

    • Vadim Gladyshev has said that long lived animals like the whale actually measure older than they are and that better systemic compensation allows their long life.

    • If they are making the organisms “older” than they are not an anti-aging intervention. You can’t have an effective anti-aging intervention that makes someone older… this is a double negative.

      • Steve – I agree this is just the kind of study we need to validate methylation clocks and develop better ones. But I think you’re overstating what can be concluded from the experiment. It is about one intervention, and they report results from only one methylation clock — a clock that’s out of the mainstream, that I never heard of.

        Many methylation sites are affected, and the fact that some combination of these in the clock algorithm showed a net reduction does not give us the detailed information we need about every site in particular, and whether changing it has a causal effect on lifespan.

  3. Aging clocks are trying to create a valid quantification of an attribute that is drowning in a sea of confirmation bias, ego reinforcement, monetary gain, etc, etc.. Don’t agree? Just watch David Sinclair announce how much younger he is (as measured by markers derived by a company in which he invested) and then sharing the nutriceutals he uses which are sold by another company he owns. …

    My background is professional sports. Derived from this field my measure of aging is what can you do via movement, strength, speed, mental agility, emotional reactions and more.

    This does not lie. You can hack it and be better than your peers but you can’t compare to a 25 year old.

    The other clocks are for ego and science studies trying to quantify a profound intangible.

    • imo the clocks can be worked out as a reasonably valid predictor, but they are not there atm… and may take some time. The rise of Ai may help with these mass data type problems, I hope.

  4. just maybe smoking, beyond demaging DNA, also is demaging methylation patterns too. But, intuitevely this alone would be stochastic, and it won’t show in positive correlation in clock. But the demaging elements are attacking certain points, and those points only. Smoke and tobacco are targeted toxins. It was the case on the DNA. It seems it is the case on epigenetic marks too.

  5. agree 100% this has always been the problem with the clocks, the clocks will not be accurate until we have a solid understanding of the role played by the genes that are being unmethylated or methylated, we cannot assume that the methylation status of any given gene has a pro or anti life expectancy effect on the organism until we know its role and how it relates to life expectancy. Given how complex and counter intuitive so many of the roles or products of these genes are and their expression this is much more complex than the methylation clocks are often made out… imo.

  6. What you write is ineresting and surely correct to a degree. However, if an intervention treats a certain type of aging damage, should the protective epigenetic changes for that type of damage not be dialed down?

    • Yes, exactly so. We see in our results an epigenetic change indicates (for example) less DNA repair. Does that mean that our intervention has improved integrity of DNA, therefore there is no need for repair? Or does it mean that the damage is there, but the body is neglecting to repair it? This is a crucial question. To ask this question requires cleverly-designed experiments and deep understanding of biological mechanisms. Statistics isn’t enough.

  7. If a central clock of some sort exists, how does it fit into the neo-Darwinism theory? Would it be easier for us to find it if we assumed methylation’s a programmed process to fight damage? Or could it be methylation itself, because it probably takes away body’s resources? Then what’s methylation driven by?

    • Methylation doesn’t cost the body anything substantial.
      Neo-Darwinists agree that there is a clock that controls development. Some would say that there’s a bug in the program that is responsible for aging, beginning after development and sexual maturation are finished. I don’t think this has much credibility, although I respect Dr Blagosklonny who has done the most to articulate this position.

      It doesn’t make much sense to me that damage accumulates and there is an evolutionary program that knows that damage will accumulate, and therefore repair is dialed up late in life. If there is evolutionary pressure to repair the damage, it would result in nipping the damage in the bud, so it never took hold, rather than ramping up the repair mechanisms in our 80s and 90s, long after the damage is done.

      For this reason, I theorize that beneficial epigenetic changes that come late in life are not on a timer, but are a response to perceived damage.

  8. A great story. Incidentally I was just looking for a reference to support the idea of “type 1” and “type 2” epigenetics in a paper I am writing. De mortuis nil nisi bonum, but of course Skulachev’s ideas of programmed aging are too clearly pseudo-science to bother refute them. But – I applaud the choice of venue! Let’s see if they publish it.

    • I knew Skulachev personally, and though I saw him rarely, we had a fondness for one another. His journal invites my contribution each year, and there is no question of it being accepted for publication. No other journal welcomes my submissions in this way.

  9. what is happening in yeast or any other unicellular organisms with predominantly asexual reproduction (https://www.nature.com/articles/nrg1706) is totally irrelevant. Of course in a local population of yeast produced by mitosis and thus consisting of identical cells eliminating an old, damage-accumulated cell is going to be evolutionarily supported. Just like apoptosis of damaged cells in a multicellular organisms. In a sexually reproducing multicellular organisms any altruistic aging strategy will be helpless against an invasion of a non-aging cheater.

    • The evolutionary establishment disagreed and fought like hell to keep Longo’s early work from being published. The idea that apoptosis could exist in a single-celled organism defies the paradigm of the selfish gene, so it was considered impossible.

  10. Well in a not to many few years would should start having data from Dr. Greg Fahy’s three drug cocktails that role back apparent age via epigenetic clocks. By that I mean data on actual changes (if any) in life expectancy.

    “Curing all autoimmune diseases/ Dr Greg Fahy”

    https://www.youtube.com/watch?v=jyanfx05cGY&t=529s

    In the follow-up version TRIIM-XA he (Fahy) made reference to his test group being several years older (looking at his graph ave. age ~67yrs) than the first group and somewhat larger (50-60) as opposed to the original nine. Sorry but if you start treating people north of 65 yrs old (especially males) you will very quickly lets say within a decade (or less) have hard data on whether or not life expectancy is actually increasing /decreasing/ or remaining the same. As well as health span. We would only have to wait not to terribly long to have such results good or bad.

  11. would you care to elaborate on why this is unlikely?

    Central to my understanding is that certain types of damages might infer information of the environment to the body.

    Sensing damage and readjusting damage levels to re-align with the ‘correct’ epigenetic program for the new environment seems fine then. Including extra damage.

  12. I got past the recaptcha by doing the following. A new hard drive, reinstalling the latest Mint Linux version – 21.3, then installing the chromium browser, and finally running all updates for both Chromium and Mint Linux.

  13. Josh, there were a couple of studies we heard about from Yuvan: the hand study, the dog study, and the new rat study with the 45 day treatment cycle. I think at least the dog trial is fully funded. Did you hear when they might be reporting the first round of incremental data for any of them?

      • It looked like a maybe, but that at the very least I think Yuvan is still working on it. Akshay and Harold seem to have two different approaches to operations that result in two very different timelines, and I think that is source of the rift. It seemed like Harold wanted perhaps academic or some other resources to do another parallel trial to allow him to move faster, but it wasn’t clear to me that he was leaving Yuvan or not. It looks like with enough money they could have it both ways unless there is some concerns about the repercussions of moving faster. Josh hasn’t posted much on this so the information sources are of course limited.

        I think if more information from new trials report it could help with Josh’s work in this paper too, since the length of life extension in the trials didn’t reveal as much about the type of epigenetic changes that mattered, but if the sample size had been bigger or the life extension even farther (and we got Dr.Fahy’s results from analyzing the frozen Sprague Dawley specimens) we could have perhaps make more definitive inferences on out the influence of both factors.

        • Second posting attempt. Adam, I have been in direct contact with Akshay. He has been stymied by the sites recaptcha. I finally got through by the technique given just above in the thread. I was told the following – Dr Katcher has submitted his paper containing the post-mortem information several weeks ago to his preferred journal. No word back yet, but this takes time (peer review, ect.). Akshay said he will send me information on the progress in a month or two. I will post it here once I get it. We all (including me) need patience.

          • Thank You George. It’s very much appreciated. I’m having trouble posting too, but it never asks me to actually complete the captcha. I also never get the emails and I check my spam folder so I just check in every so often looking for updates.

            You might try a VirtualBox image and refresh it when you run into trouble so you don’t have to reinstall your entire system each time.

            Josh, did you see this? ‘Causality-enriched epigenetic age uncouples damage and adaptation’
            https://www.nature.com/articles/s43587-023-00557-0

          • Harold’s Linkedin account says he has left the company and is seeking a lead researcher role elsewhere.

Leave a Reply

Your email address will not be published. Required fields are marked *