In her lab at Yale, Dr. Morgan Levine tackles some of the most exciting and difficult problems in geroscience. She specializes in bioinformatics and is working on creating and finessing clocks that measure biological age. We talked with Dr. Levine about her work, the impact that biological age clocks have had on the longevity field, and how due diligence is needed when applying the powerful methods of machine learning to biology research.
This year, we mark 10 years since the creation of the first epigenetic (methylation) clock . How has this invention advanced the longevity field, where do we stand today, and what are the next steps?
Epigenetic clocks hold a lot of promise for the field. One of the problems in aging research is that biological aging is a latent variable – you can’t actually observe it. Therefore, people had to find proxies for testing whether specific interventions affect aging. Lifespan has been used in most studies, but now more research is being done in humans. This requires more immediate outputs that can determine an effect on biological aging.
The epigenetic clocks are one way to do this, but I still don’t feel we have a biomarker of aging that is perfect and has no caveats. There’s a lot of room for improvement in the epigenetic clocks field, and I’m excited that we’re going in that direction, whether it’s understanding mechanisms or what the clocks are actually capturing.
It is important to understand which kinds of methylation changes are critical for lifespan and healthspan, because you can change the clock, but it doesn’t mean you changed the important components that will actually translate into a longer healthier life. Disentangling that part will be critical. And then getting them to be really robust, like N-of-1, on an individual level – that would be awesome.
Can we say that the clocks have revolutionized our field, or would that be an overstatement?
I think people are now seeing the benefits of having biomarkers of aging in real research (and I’m not saying epigenetic clocks are the best ones). Epigenetic clocks were the first to be taken seriously as something that can be used to proxy aging. While many of their critics have valid points, it’s an amazing first step to show that we can do this. Maybe not perfectly, but we’re getting there.
You have made a breakthrough with the new clock that, as of July, didn’t even have a name. Does it now, and could you walk our readers through that?
So, this clock is technically not published yet, and we’re still improving it to make it even better, but right now, it works pretty well. At least internally, we’re calling it SystemsAge clock, but it still doesn’t have an agreed-upon name. Once we actually publish it, we’ll have to come up with a final name. I’m not as creative with naming as Steve [Horvath] was with his GrimAge clock.
What we tried to do was somewhat similar to the PhenoAge or GrimAge clocks – to use clinical chemistry or proteomic measures as something to train on. But in this case, we used a priori evidence of which organ systems different biomarkers reflect.
We grouped the biomarkers into a bunch of different systems: inflammation and immune functioning, kidney functioning, liver, etc. And then, like with PhenoAge, we actually made aging measures of that and trained methylation predictors on each of the systems.
Then we combined them into a full systems clock, which captures the combination of all these things. But even more exciting is that we can cluster people based on their profiles across systems. Two people can have the same epigenetic age and even the same chronological age, but they could have gotten there through different routes. One might have chronic inflammation, and another person might have impaired liver functioning. We want to understand those individual profiles, what’s associated with them, and what those profiles mean for different outcomes. That’s where we are right now.
In another paper, your group was able to drastically lower the sample sizes needed to detect effects of interventions .
Yes, that was us redoing existing clocks. We found that almost all the existing epigenetic clocks in the literature are very noisy. Say, I take a blood sample from someone, split it in two and run it twice. As a result, I can get differences of up to nine years in the predicted age – and that’s in the same batch, run on the same array. So, it’s not like a weird batch effect, it’s just a noise issue.
We realized that to be used in a clinical trial, or even at an N-of-1 for people who are interested in biohacking or quantified self, we have to be very reliable at least in the numbers that people get. So, we employed a new kind of statistical method that we developed for making these clocks where instead of training clocks based on DNA methylation values from CpGs [methylation sites in the genome where cytosine is followed by guanine], we first ran a principal component analysis.
We used essentially all the principal components as input to the elastic net, the same way that CpGs used to be input. The nice thing about that is that random technical variation will not be captured in the principal components. They won’t pick that up. This way, you almost completely remove this kind of random or stochastic signal from your measures and you just get whatever the non-stochastic signals or functions are in your data.
We found that now, we can get most of the technical replicates within a year of biological age of each other. Then we made new versions of the most used clocks like the original Horvath clock, the skin-blood Horvath clock, GrimAge, PhenoAge, Hannum – basically most of those that are widely applied.
And we showed that this can increase the power of clinical trials. If you’re expecting your clinical trial to produce a difference of, say, two years of change in epigenetic age, because we’ve taken out the noise component, you actually don’t need as many samples to find that.
This sounds huge. Do you think it will be used by other people in their clocks?
Yes, we’re already having people contact us, and we’re putting up a GitHub, so people can actually download the code. We want people to use it. And before that’s up, if people email us, we’re also trying to make sure we can distribute the code so that they can just calculate these clocks in their own data.
Going back for a second to your new clock. So, it is superior, at least in some parameters, to the existing clocks?
We are trying to test a number of different outcomes. In terms of just overall mortality prediction, it’s doing better than not just GrimAge, but also this new cleaned-up version of it that we’ve made. It also is much better in measuring functional aging outcomes, like cognitive functioning, physical functioning, etc. We have found that GrimAge is very strongly driven by cardiovascular diseases, and it’s not as good in predicting things that are a little separated from that. So, our new clock does better in cardiovascular mortality and way better in these other aging domains as well.
Methylation clocks have always been sort of a black box. We know that their readings correlate with aging (even if we may not agree on the definition of aging), but we don’t know how they do that, what the biological underpinnings are. How can we solve that?
That’s something we’re most interested in – figuring out what’s under the hood, what is actually driving these measures. And our hypothesis is that there’s not a singular phenomenon, there are different types of methylation changes throughout the genome that represent different biological phenomena or pathways. They map to different things, perhaps different hallmarks. We are trying to deconstruct epigenetic clocks to really understand which age-related methylation changes track together across tissues and samples.
And from there, we’ll be able to do a better job at connecting them to some functional or mechanistic underpinnings. Whereas if you look at the clock as a whole, it’s kind of a grab bag of many things that are contributing to it. We’re also doing a lot of in vitro work, both in human and in mouse cells, where we can induce things and see the changes.
Is it possible that the abundance of epigenetic clocks leads to some researchers cherry-picking them to find those that will yield the best result?
I think that probably happens. My guess is that people will test a few clocks and then they put out the one that worked. And I’m not saying I know of anyone specifically doing that, but I do think it’s definitely not just possible, but it’s probable that people are going to do that because there are so many options now.
What we like to do, and what I like to tell others who I collaborate with: test multiple clocks, but – and I think it’s important – put them all out because knowing how things map to different clocks is informative, especially with the second-generation clocks, which were trained not on chronological age, but on biological aging correlates. That gets us back to this whole idea of deconstructing the clocks, understanding why certain clocks work in certain contexts and others don’t.
The unwillingness to publish results that are less than flattering is indeed a major problem.
Yes. If only people just published their data. That’s what I would like to see: just publish the data and anyone should be able to test different clocks on it — including new clocks that come along.
You have found that reprogramming drastically reduces cellular biological age. This is a hot topic, with Altos Labs now joining Calico and others. What are your thoughts on cellular reprogramming in the context of aging?
Reprogramming is something we’re extremely interested in. We are not approaching it necessarily with the idea that it’s translational at this point. We’re interested in it as a fundamental biological question. And it seems to be very much dictated by the epigenome. We want to know what epigenetic patterns dictate cellular state. How do cells traverse this landscape and end up in a different cellular state, and what’s dictating that? Do cells directly go back up the same path they came down? What would it mean if you age that cell again in vitro? We’re very curious about all these questions.
What does your scientific intuition tell you? Could this be the Holy Grail, the thing that will end aging?
I don’t know if it will end aging. I like to think of aging the same way you would think of an ecosystem. The living systems are important for the ecosystem, but there also are non-living systems – in our case, all these byproducts in the microenvironment.
Such as the intracellular matrix?
Yes, the dynamics between both the living entities and their environment are critical. I think reprogramming is probably not going to solve aging on its own, but it is a very exciting scientific question. I think that potentially we can figure out feasibility and how to do it safely. We can start from a small scale. If you just intervene in one cell type, does that get you any benefit – and that’s without trying to completely move the entire organism back to some stage.
Are there any actionable strategies that are known to decrease epigenetic age, such as caloric restriction?
In mouse studies, caloric restriction repeatedly seems to reduce epigenetic age across different clocks in various labs. I would say, that and reprogramming are the two things I’m very confident about – in mice – both at least make it look like the epigenetic clock has reversed.
In humans, it’s less clear. Definitely, cellular reprogramming still shows this. With dietary restriction studies, we’re not sure yet, but I hope that they would, and I think they might. More research needs to be done on fasting and on different dietary restriction regimens other than just straightforward caloric restriction.
Yes, that would be very interesting. And is there something that you personally do to, so to speak, keep aging at bay?
I eat almost exclusively plant-based food and I do time restriction eating like a lot of people do. I also try to get a lot of exercise. My sleep is one area I’m not sure I do well in.
I agree, that’s hard. Well, let us look at the broader picture. How important are bioinformatics to medicine as a whole and to the longevity field in particular?
They can be critically important, but we shouldn’t assume that all applications of them are perfect and somehow the magic answer to everything. From people outside of bioinformatics and computational biology, there’s too much leeway or optimism about what certain models actually can provide. It’s like “garbage in – garbage out”. If they’re applied in a thoughtful, creative way, they can be very powerful, especially with aging, which is such a complex process with so many different variables that are changing. The only way to truly capture that is through approaches based on computational biology or informatics. But again, not every model is useful.
Yes, we are seeing a lot of machine learning in research lately, and there’s a feeling that these tools sometimes are used without proper training or understanding. Do you think this might be spreading a bit too fast?
Yes. The fact is that it’s become a lot easier for people who don’t really understand what the tools are doing or how to creatively use them together, or what assumptions you’re making in your data. Anyone can just apply it. I can give undergrads or brand-new graduate students in my lab some R or Python library, give them some script, and they can put new data in and run it. But that doesn’t mean it’s a good model.
The other thing I find that often is lacking is creativity in terms of how we’re doing this. We need to think more about what question we’re trying to answer and what assumptions we’re making, and then figure out creatively how to do that. It’s not enough to throw something into a neural net and say, oh, see, I can get some predictions, so it must be useful.
Maybe people like you can do something to educate the scientific community on how to use these tools properly?
I hope so. It’s kind of what you alluded to – that people can apply these tools really quickly, but it takes some level of intuition about the data to be able to actually use them in a smart way. That only comes from doing a lot of data analysis for a long time and knowing what data looks like and what things to look out for. So, yes, I think we can talk about it openly, so that the general community doesn’t just accept that every single thing that comes out of bioinformatics is gold.
A couple of words on your collaboration with Elysium. You have developed an at-home epigenetic age test for them. Could you please tell our readers more about that?
I actually have a book – like everyone else – coming out. It will be published next May, and this is basically the topic of the book: how can we measure aging, track aging, quantify it? How can that be used by people and what can you eventually do about it?
Elysium approached me right after I started my position at Yale or maybe a year in. My understanding is that they were selling NAD precursors that they suggested would slow aging, but they had no way to verify that in humans. So, they were looking for biomarkers of aging. It was shortly after the PhenoAge paper had come out. We developed this collaboration where at first, they were licensing a different version of PhenoAge, but then we discovered this issue with reproducibility where you can measure the same sample multiple times and get a different answer. For a consumer-facing product this is a huge problem. So, I got kind of obsessed with solving this problem and eventually I figured out how to do this. I developed another clock, which is what they eventually licensed and now provide for this consumer-based tests. I’m serving more as an advisor but not doing a lot of hands-on work with them besides providing ideas, answering questions about how things can be applied.
Just to get the idea: how do results look for the consumer?
You get a kit in the mail with a saliva test, you send it back, and then you get an email that tells you your epigenetic age, your rate of aging, which is your epigenetic age divided by your chronological age. A difference of two years is, obviously, a bigger deal when you’re 20 versus when you’re 80.
And then they just give you a straightforward lifestyle recommendation sheet based on observational studies. We actually don’t know anything causal, it’s just epidemiological data, and it’s nothing people probably don’t already know – like don’t smoke, eat plant-based, exercise.
So, the added value is that it can alert people of their condition and motivate them to change something in their lifestyles. And they also offer a subscription because the longitudinal aspect is important, I guess?
Yes. I actually don’t know a lot because I’m not really involved in any of the business stuff. I believe they offer that. As a scientist I would say that the longitudinal aspect is important, assuming you can afford it, and the question is how often someone should be measuring. It is important because I don’t necessarily care what my epigenetic age is today, I just care that over the next 10 years, I’ve only increased it by less than 10 years. You want to keep that rate slower.
One last question: are you familiar with any research into the methylation patterns of centenarians and supercentenarians? Do you think it’s an important thing to study?
The hard thing with doing research in centenarians and supercentenarians is there’s no control group to compare them against. All the individuals who had died – you don’t know what they would look like if somehow you kept them alive. What has been done, including in some of the studies I’ve been on, is comparing offspring of centenarians to control. And in these age-matched controls, we do find lower epigenetic age using mostly the second-generation clocks. But this is the issue with looking at centenarians: let’s say we find that, on average, they’re five years or even 10 years younger than their chronological age. The problem is that epigenetic clocks don’t increase very linearly with chronological age. And we do see a slowing or at least a decrease, even starting at age 70 or 80. So we don’t know if they are along the trajectory that would have continued, or if they are truly younger.
 Bocklandt, S., Lin, W., Sehl, M. E., Sánchez, F. J., Sinsheimer, J. S., Horvath, S., & Vilain, E. (2011). Epigenetic predictor of age. PloS one, 6(6), e14821.
 Higgins-Chen, A. T., Thrush, K. L., Wang, Y., Kuo, P. L., Wang, M., Minteer, C. J., … & Levine, M. E. (2021). A computational solution for bolstering reliability of epigenetic clocks: Implications for clinical trials and longitudinal tracking. bioRxiv.