#386 - Aging clocks—what they measure, how they work, and their clinical and real-world relevance

>> Hey everyone, welcome to the Drive Podcast.

I'm your host, Peter Adia. This podcast, my website, and my weekly newsletter, all focus on the goal of translating the science of longevity into something accessible for everyone. Our goal is to provide the best content and health and wellness

and we've established a great team of analysts to make this happen. It is extremely important to me to provide all of this content without relying on paid ads. To do this, our work is made entirely possible by our members. And in return, we offer exclusive member-only content and benefits

above and beyond what is available for free.

“If you want to take your knowledge of this space to the next level,”

it's our goal to ensure members get back much more than the price of this subscription. If you want to learn more about the benefits for premium membership, head over to peteratiamed.com/subscribe. [MUSIC]

Welcome to a special episode of The Drive.

In this episode, I take a different approach where I walk through a single topic in depth. And this is a topic that many of you have been asking about. Agent clocks. The way this episode, I explain what Agent clocks are

and the difference between chronological age and biological age along with the difference between those and something called the pace of aging. How epigenetic clocks work and what they may actually be measuring. I'm going to talk about a randomized control trial

that used three very simple interventions and tested four of the most common aging clocks. I'm going to also talk about another study that used brain imaging via MRI to study the pace of aging

“and see what could be gleaned about not just the risk of dementia,”

but also mortality. I'll discuss the biggest limitation in the field, which is whether changing a clock actually changes meaningful clinical outcomes. So without further delay, I hope you enjoy this special episode

of The Drive. [MUSIC] So if you wanted to run the perfect anti-aging trial, the endpoints would be really obvious. You'd want to see fewer heart attacks, fewer cancers,

fewer dementia diagnoses and ultimately fewer deaths.

So you would call these hard outcomes real outcomes that matter. These are the clinical outcomes that we all care about. Now, of course, the reason we don't see these trials is that they would take a very long time.

These would literally be 20 year trials and with that would come in order enormous complexity and cost. Further, more it would be very difficult to ensure that whatever intervention you put in place was being put in place for the duration of this time.

I mean, that would be not that hard to do if it was a drug trial because it's relatively easy to take a drug, but it would be more challenging for a lifestyle trial. Okay. So every few years, the fields of neuroscience and medicine

and cardiovascular diseases, etc., they go looking for a proxy or a shortcut. So some intermediate marker that could move faster than these hard outcomes, but that would still predict the hard outcome reliably.

And I think over the past few years, what we've really seen is that aging clocks are the most interesting and popular proposed shortcut. Again, I don't use the word shortcut with a sort of negative connotation.

It's like, this is what we need. We do need a shortcut. We need a proxy. So again, the idea here's pretty compelling. Imagine you could have a single number

that would predict your actual aging or your actual biological age that's different from your chronological age or maybe a rate of aging that reflects a new trajectory around this could be very valuable in designing clinical trials

or looking at interventions and even at the individual level understanding if you've made a change and is it making a difference? So this would think about this as a foray

into precision medicine. Now, there's a little bit of a problem in my mind because these aging clocks are being marketed as the latest and maybe best way to keep tabs on your health. Lots of people are ordering them.

They're available to anybody. They're sold by longevity docs who promise to improve your biologic age with groundbreaking combinations of peptides or other elixirs.

“But I think it's worth looking into these”

a little more closely to understand what the science can actually tell us. And I think the best way to do this is to look closely at two studies to very interesting studies

that can help us get at the fundamental questions that we really want to be asking around this, which is what is the clinical utility of an aging clock?

I just want to kind of make sure everybody's starting from the same footing in terms of understanding the biological stuff that we're talking about. So what is an aging clock?

We'll basically edit its core.

It's a prediction model. But let's take a step back. Your chronological age is also a prediction model. You see, if I told you that in front of me there is a 20-year-old and there is a 70-year-old

and I asked you to predict which one of those people is going to die first.

“I think everybody would knowing nothing else”

make the correct prediction. Now we could layer onto that certain other factors. So if I said, "Okay, well I actually now have two 70-year-olds in front of me

and one of them has cancer and the other one does not. Do you predict which one of those is going to live longer than the other?" And again, without knowing anything beyond what I told you,

I think everybody would make the same prediction.

And so this idea of using information

to predict mortality is not new. It is the entire basis of the actuarial underwriting industry. And there were companies that are exceptionally good at doing this.

These are called life insurance companies. And their data are incredibly proprietary. And it's really less so their data and more so what they do with the data that is incredibly proprietary.

They gather a lot of information about you. They do a blood draw on you. They know your age. They know various factors about you. They take your blood pressure

or your weight and things like that. Relatively rudimentary stuff. But from that, they have these tables. Again, highly proprietary that seem to do a very good job of predicting

when you're going to die. And so the question is, would one of these aging clocks be even better? Okay, so let's talk about how these things work. So they typically work by starting with some biological data.

And the most common thing that we're going to hear about is epigenetic data. So this is DNA methylation. And then they train an algorithm to look at that and predict something age related.

So I think it's worth spending a minute on DNA methylation. I don't want to go far down the rabbit hole on this. But you've undoubtedly heard the term. And so I just want to make sure everybody's playing

the same level. So DNA methylation is a way that we, with the body, modifies epigenetically what the DNA expression is. So it doesn't change the sequence of DNA, but it can influence how the genes

are turned on or turned off.

“So that's what we call expression of genes.”

So when you modify the epigenome,

which is basically when you put a methyl group,

so that's a carbon with three hydrogens, when you put it on the backbone of the DNA, that impacts whether or not that section of DNA gets turned into RNA. That's what we mean by expression.

You may have heard the term CPG, but not in reference to consumer package goods, but a CPG refers to the location where these methylation's most commonly take place. If you remember back to high school biology,

we have these four nucleotides. The C is the abbreviation for cytosine, and so where these things typically occur is right on the phosphate bond that links the C, the cytosine,

with the G, the guanning nucleotide. So when we talk about CPGs, that's just kind of another way that people kind of quickly talk about the methylation. So the methylation that occur at these CPG sites

will then affect quite strongly the genes that are near to those areas. And why we care about this, of course why I'm even talking about this, is these methylation levels of many sites

actually change somewhat predictably as we age.

“So this is kind of the rationale for all of this, right?”

Is as we age methylation sites change, air go, if we can measure what's happening at methylation sites, can we impute age? Can we impute something better than chronologic age? Because remember, chronological age

is an awesome predictor as it stands. But we're asking, can we do better? Because chronologic age is great at telling you that on average a 60 year old is going to live a shorter duration than a 50 year old.

But we know that that's not true at the individual level. There are plenty of 50 year olds that are going to have a shorter remaining life than plenty of 60 year olds.

Just depends on the individual health and a whole bunch of other things. So we want to get it that different. Okay. So these patterns are going to shift gradually over time.

And various factors, behavioral factors such as smoking,

actually play a role in that. And so for this reason, researchers came to the conclusion, roughly 10, little over 10 years ago, that we could use these as kind of a molecular

record keeping of what's going on in the body. And I think that's probably why DNA methylation has created such an important and foundational part of the clock story.

“So that's why I kind of went a little deep”

in the weeds there. I think it's important. So, you know, you've probably heard of the Horvath clock.

That's that's one of the earliest first

generation clocks. And that was obviously based on this. So these models were trained on large data sets of DNA methylation, which were measured and collected

from thousands of individuals across a wide age range. And they're mostly using cross-sectional cohorts rather than tracking an individual over time. Why? Because is exciting as it would be to track an individual over time.

Those data sets are somewhat limited harder to get. Whereas if you take very large cohorts, where you just slice the population, you would get access to 20 year olds, 25 year olds, 30 year olds,

50 year olds, 60 year olds, 70 year olds, 90 year olds, etc. And the hope would be that hey, we're going to see what the signature of methylation is, you know, over time. So given that chronological age

was the outcome that the model was trying to predict. It's not really surprising that these clocks became very good at estimating age, often within kind of a few years of age. But at the same time, the fact that patterns of DNA methylation

change consistently across ages of individuals that had a biological interest to it. And it suggested that there might be certain areas in the genome where methylation shifts occur in a predictable way as people get older. Okay.

“So from a clinical standpoint, what does that tell us?”

Well, estimating chronological age,

which is what these first generation clocks did,

wasn't adding any value because we already know chronological age. So it was more of a proof of concept, but that's when the researchers realized, okay, what we really want to do is come up with something that could be better than chronological age, and that's where this idea of biological age could come from.

And I just want to explain like sort of an extreme example of what this would look like. So again, let's say you took two 60 year olds, who, I didn't tell you anything else about them other than they're both 60, and let's just say they're both the same sex. So two 60 year old women.

They're an actuarial table, would say based on their chronological age, these women both would have a life expectancy of, I'm making this up, 27 years. So if you're 60, your life expectancy is 27 years,

you're expected to live to 87. But if we could look at the methylation of these two women, and one of them came back and we were told yes, but her biological age is 65, and the other one's biological age is 55.

The question is, does that delta of 10 years between them, actually translate to 10 years difference in lifespan?

“And so that's what these next generation of clocks set out to do.”

Instead of predicting age, they started training on more clinically, meaningful outcomes. So they were trained on physiological biomarkers, data sets that would be able to track mortality,

and basically even things like rate of physiologic decline,

or pace of aging. So the last category, this pace of aging, when I think is particularly interesting, because it aligns with what most people think they're getting, when they look at a biological clock, which is not just how old am I,

but how quickly am I aging right now? This topic came up, by the way, in a podcast, with Rich Miller, which was when they went back and looked at data from the various ITP winners, or recall Rich Miller is the guy who oversees one of the parts of the interventions testing program at the University of Michigan,

where they take drugs like rapid mice in metformin, nicotine in my riboside, et cetera. They put them into mice over the duration of their life, and then they look to see which of these drugs extend life. Well, they took all of the ITP winners,

plus other things that were known to improve lifespan of mice, such as coloric restriction, and they identified roughly a dozen, or so, not epigenetic changes, but actual things that showed up proteins that showed up in those animals,

something we would prefer to as the proteome, and they were able to generate in those mice predictive age rate calculators. So again, a little too soon to tell if that's going to pan out in humans, but very interesting. So many of these second generation clocks were designed to predict outcomes

like mortality by incorporating methylation patterns that correlated with things like

and while that's useful, it's important to understand that that kind of

changes how we interpret the output. So if a clock is particularly capturing a biological fingerprint of something like smoking history or cardiometabolic health, then a shift in that clock might reflect an improvement in that pathway, but it might reflect something really uninteresting, like you're just recovering from a cold,

or you have lingering inflammation from a heavy workout. So this is kind of where the excitement around aging clocks collides with the reality

“of measurements, because there's basically two types of noise you have to consider”

when you look at these clocks. You have to consider biological noise, which is the example I just gave of how do I know that what you did in the day or week before didn't transiently impact what I'm measuring, but truly has no impact on your health versus the measurement noise, which is how hard is it to actually measure these things? So even in a very high quality lab,

DNA methylation measurements are not very stable. So variations can arise from a lot of things that how the sample is handled and stored, differences in the methods for DNA extraction, you know, the efficiency of the conversion steps that are used to read the methylation patterns, the batch effects on methylation, arrays themselves, and even differences in the mixture

of immune cells present in the blood at the time that the blood is drawn, can remember these are all being done not on tissue but on cells in the blood. And then on top of that, you know, the clocks are typically measuring hundreds of thousands of methylation sites across the genome, and then trying to collapse all of that information into a single summary score.

And again, we love when we can do that when we can convert lots of data into a number, but we have to understand that we run risks of doing that. So again, I'm not saying any of these things aren't challenges that maybe couldn't be overcome, but I just want to make sure everybody understands how technically complicated this is. So again, just keep in mind biological noise and then technical or measurement noise.

So basically, I think the reason these clocks are exciting is that they're kind of offering

three things that people want. So the first one I just kind of alluded to, which is compression. So aging is very multidimensional. There are so many things that are going on. Your immune system is declining. Your fuel partitioning skills are declining. Your health is declining. Brain functions. All these things are declining as we age on average,

but the clocks are attempting to take this and compress it into a single number. On the one hand, that's exciting. We love when we can do things like that. BMI is a great example. Again, it's only combining height and weight, but it's turning it into a single number. But we have to be mindful of the fact that the more you compress something complicated

and BMI is a great example. It's only taking height and weight and compressing it into a single number and trying to be a proxy for muscle mass body fat and things of those nature.

“And here's the thing. On average, it's pretty good, right?”

That the population level BMI is pretty good. If I told you that in one city, the BMI average was 24, and in the other city, the BMI average was 29, and I said, tell me which of those two cities you think is healthier. Again, knowing nothing else, you're all going to say 24,

and you're almost assuredly right. But if I told you I have an individual who's got to BMI a 24, and I have an individual who's BMI is 29, tell me which one's healthier. Truth of the matter is, it's going to be tough. There are some really unhealthy people with normal BMIs,

and there are some really healthy people, usually quite muscular, that have quite high BMIs. So this idea sort of falls apart and you can imagine how much more susceptible we would be to that when we were taking a much larger and more multi-dimensional problem. Okay, the second big thing we want out of clocks is not just compression,

“but it's speed. I kind of alluded to this at the outset, right?”

It would be amazing if we could do clinical trials for a year

and get the type of benefits in terms of insight that we would get if we were doing these clinical trials for 20 years. The third one I also kind of alluded to already, which is at the individual level, we want feedback. I want to know that if I change my diet from this to that,

after three months or six months, was that the right change to make. I want to know that if I'm taking this supplement, which supposedly reduces inflammation and supposedly does several other factors that are targeting hallmarks of aging, I want to know if it's doing it.

So bottom line is, this is why we want them, question is, does it work? Okay, so the best way we could think to talk about this is

that we thought really highlighted a couple of the important points here,

“and so that's what I want to talk about here are these two studies.”

Okay, so the first study looked at an intervention.

They used omega-3 supplementation, vitamin D supplementation, and exercise. And then they asked the question, will those simple interventions, and I'll talk about them in a little bit more detail, move the needle on the four most common epigenetic aging clocks in a randomized fashion. The randomized people to these things much longer.

Second study that we're going to talk about takes a different approach, but asks if we can estimate a person's pace of aging, using structural features from a single brain MRI. So different approach, but let's talk about them both. Okay, so let's start with the first study.

This was referred to as the do-health study, and I think it's a reasonable use case example. Okay, so again, if an agent clock is going to be useful,

“the best case scenario is probably that it can detect biological signals”

and a randomized control trial. So if you think back to, again, the problem we started with, the endpoint you really care about is something like mortality, or incidents of disease like dementia, cardiovascular disease, or cancer. And the interventions in this trial, which are very reasonable things to propose,

you're not going to figure out in a year, two years or three years if they're having an impact on those things, but if you have a biological clock, that is true to those things, you're going to be in the end zone. So what did the study do? So the study took, you know, it was a European large study.

It was a two by two by two factorial. So that means they tested these three interventions that I talked about, so giving vitamin D, giving EPA, DHA, and assigning exercise. They tested them individually and in combination. So each individual is in randomized to one of eight groups for the duration of the study,

which is a three year duration study. The measurements were collected at baseline and then at three years. So you got a blood level at time zero and then at three years. So they looked at nearly 800 generally healthy, older adults. So everyone was 70 plus, mean age 75 from Switzerland.

About half of these individuals met the criteria for what we would call healthy aging.

So they were basically free of chronic diseases, disabilities, cognitive impairment,

any other limitations. And they were quite active about 88% of these people reported regular physical activity. And about 60% of these people reported exercising more than three days per week prior to enrollment. Okay. So again, the interventions, 2000 IU of vitamin D, a relatively modest amount of EPA and DHA.

So one gram a day that contained 330 milligrams of EPA, 660 milligrams of DHA. And this was from marine algae. And then adherence for both the vitamin D and omega three levels was assessed by changes in serum level. So they did have the ability to kind of go up or down based on what they were measuring. And then on the activity front, they had a simple home-based exercise regimen that consisted

of mostly strength training 30 minutes, 3 times a week. And this was added on top of whatever you were doing at baseline. So whatever you were doing at baseline fine, but we add this. And then compliance was tracked with exercise diaries and follow-up.

So my first thought when I looked at this was these are kind of modest interventions.

You know, a gram of omega three is pretty low. 2000 IU of vitamin D is not going to do much in 90 minutes of exercise. In people who are mostly already exercising, you know, I guess it depends on what they're already doing, but truthfully I would have thought that was the most interesting thing,

but none of these are herkially in things. Especially when you combine it with the fact that most of these participants were relatively healthy and physically active. So it's not clear what you would see here.

“But that said, I think this trial was designed well,”

and it's a useful way to test whether these aging clocks can detect subtle changes over time. So let's talk about the test, right? So again, time zero and time three years, they measured DNA methylation, and then they applied these clocks. So let's talk about the clocks that they used,

because I mentioned that they used four next gen clocks. So the first one is called phenoage. So this test uses methylation data from about 500 CPG sites, and it's trained to reproduce a clinical biomarker score that predicts mortality risk. So that biomarker's more than incorporate takes measurements beyond just the CPGs.

It looks like it looks at albumin, glucose, C-reactive protein, kidney function, white blood cell, count, and then the clock reflects the physiologic health,

The next one is called Grimage. This uses methylation data from about 1,000 CPG sites to estimate the level of plasma proteins that are linked to aging, such as GDF, 15, leptin, P-A-I-1, and also smoking exposures. So these methylation derived biomarkers are then combined with bi-- our sort of chronological agent, sex, to predict a time to death.

So this is kind of an interesting one, right? This is that one that would try to get at what I was saying earlier, which is if you have a 60-year-old, the actuarial time to death expectation might be, you know, I've made an up and number, but say 35 or 37 years.

If you did a Grimage on that person, and it said 20 years, that would suggest you to this person is much less healthy than the average 60 year old. And if it said, "No, 40 years," you would say, "Oh, this person's healthier than the average person that you would expect to see that age." Okay, then there's another one called Grimage 2, which is the same as Grimage. It just has an updated set of biomarkers, so it's using C-reactive protein and A-1C,

along with some other refinements. And then you have another one called the Dunedin-Pace estimate, which is really trying to estimate the rate of aging rather than biological age. So this uses the methylation patterns of 173 CPG sites. So unlike the others, which were trained on cross-sectional areas,

this is trained on a longitudinal set of data across a population in New Zealand called Dunedin.

“And that's what allowed them to track over time what the changes were,”

and try to estimate this if you will, first, derivative.

Okay, so we'll put in the show notes a table that just kind of links to all of that, that you can sort of keep that in mind and keep coming back to it. So just to summarize that, the first three, which were the Theno age, the Grimage, the Grimage 2, those are ones that are kind of looking at biological age, and trying to see if that can be more predictive than chronological age.

And then the Dunedin-Pace is designed to estimate the pace of aging. Okay, so again, important distinction, and again, what does it matter? Right, the biological age might tell you that a person's physiology resembles that of a typical person who's older or younger, that was the example I gave, the pace of aging is trying to answer a slightly different question,

which is, at the time of my measurement, how fast is the system deteriorating? Okay, so let's talk about what the study found. We'll link to the figures in the show notes, so you can go and actually see the figures yourself,

“because I think the figures sort of tell a thousand words, right?”

But basically, they looked at what each of the four tests showed,

and they have forest plots for each of them, and they look at each of the four different forest plots for each of the interventions, combinations of, and they're about. So what I'll just call out is what was significant. So the BNO age study found significance, and effect size of 0.2, for just the omega-3 intervention, and then for the omega-3 in the vitamin D intervention,

for omega-3 and exercise, and then for all treatments combined. When you looked at the grim age by itself, just this first-gen grim age, it found no significance, no change across the board, and anything, when you looked at grim age 2, the only thing that showed significance was omega-3 versus placebo,

and when you looked at the deniedin pace,

“the only thing that showed significance was omega-3 versus placebo.”

So the bottom line here is on balance.

You know, the most consistent finding was that something about omega-3 supplementation moved the needle in at least three of the four clocks. So the only clock that it didn't move the needle in was the first-gen grim age clock. Now that said, even though the omega-3 change showed consistency in three of the four tests, the magnitude of the effect was quite small.

So if you translated into something a little more intuitive, the effect corresponded to about three months of reduced aging over three years, depending on which clock you look at. And again, I still actually think that's a larger magnitude than I would affect, but maybe if you believe these clocks the translation is for people with really, really low omega-3 intake,

you don't need much to move the needle. The vitamin D supplementation, as I said, didn't really impact the clocks at all, but again, given the dosage, I don't think that was surprising.

So it's possible that just the 2000 IU wasn't enough to get them anywhere.

And then exercise also failed to show an independent effect. Again, I think context here matters. Remember these participants were already physically active at baseline. So it might be that for a study of this size and this duration, you weren't going to see the effect unless you did it in people who were not active.

So again, not sure what to make of this.

“I think that the results of this study, you know, don't answer a ton of questions.”

These interventions are quite common.

Not sure if they were dosed correctly. I don't recall actually know that I think about it. What they pre-selected as their power analysis and other words, to pick the sample size that they picked, they had to assume a certain effect size. And it might be that the effect size was, it was a real one, but it was smaller.

That would mean it might not be clinically significant. But I suppose that the use case here is reasonable, right? Which is, were there small molecular ships over time in an otherwise well-controlled setting? Of course it also begs the question, though, which of these clocks do you believe, right? Because if you think that grim age is the right clock,

then it would say none of these things matter. If you pick thino age, you would say, gosh, everything mattered except for exercise.

“And if you pick grim age too, you would say Omega 3 was the only thing that mattered in the same with the need in pace.”

So I still, at least on a personal level, don't know that I can tell which of these clocks makes the most sense. Now, one potential advantage of using an aging clock is that it does give a shared endpoint across different interventions. So if you at least believe that there's internal consistency, you could feel maybe more comfortable that, okay, in an absolute sense, I don't know if these changes are right. But I could figure out that, hey, I'm getting more bang for my buck doing more cardio versus more resistance training or vice versa.

So the other thing I think that we don't know here is what the measurement error was. So how much technical noise there was in these, we've talked about this earlier in the podcast, not this one, but previous episodes where there are lots of folks out there that will go and buy the same multiple versions of the same test. Take a blood sample, identical, you know, single blood sample and then spread it across multiple tests and you get different results. And so, you know, there are various reasons that that could be happening. It's not clear from this paper exactly what their technical spread was.

But my guess is there's more noise in these measurements than say measuring a blood glucose where the assays much easier and much more standardized lab to lab. You know, I still think that whether these clocks are actually capturing the biology we care about well enough is I think that's an unresolved question. We can obviously ask different questions with clocks, but with most medical research, we kind of focus on very specific outcomes, right? We're going to look at muscle strength. If we're testing resistance training, we're going to look at blood pressure, if you're looking at an anti-hypertensive drug or LDL cholesterol, if you're using a lipid lowering drug.

I like what the agent clocks are trying to do because not all things are going to be measured in single parameters. And even if you do lower LDL by itself, it would be nice to know how much of an impact is that having on my overall agent.

“So I think maybe the most important thing here is that if an agent clock could allow a researcher to ask a long-term question within a shorter trial,”

that alone to me is reason enough to do this and then everything else, whether we individually can use them, that those would be fantastic. I think what I liked about this paper was that the authors didn't kind of use just a single clock that gave them the answer they were hoping for. They pre-specified four clocks, they showed the data for four clocks. We talked about the results, right? Three other four showed a benefit for Omega-3. Again, is that because they're detecting using different CPG sites, is it because they're using different biomarkers?

It's not clear. The authors point out that this type of discordance is not unfamiliar. There's a very famous study called the Callary trial. This was run out of Pennington. Many years ago, Eric Robertson, who's a previous guest on the podcast actually, was the PI for that. But the data set from Calorie has been used multiple times. This was a Callary Restriction study. And it showed that the Callary Restriction showed a reduction in pace of aging using the Deniedin pace clock,

but it didn't affect the phenoage or grimage, which were the other first-generation clocks.

So again, this is not unusual. Okay. So with all that said, let's look at the second trial.

And so you'll recall that the Deniedin pace was developed using blood-based biomarkers in the Deniedin New Zealand cohort

across a longitudinal path. So this is kind of a cousin of that.

“This is called the Deniedin Pack N.I. or at least that's how I pronounce it.”

So it's Deniedin and then it's P.A.C. N.I. And this also tracks people longitudinally, but it tracks them based on a brain MRI and then following them over time.

Again, it's going to develop a rate of aging, but rather than doing it longitudinally based on blood work, it's going to do it based on brain imaging.

So, okay. The Deniedin multidisciplinary health and development study, and I just want to give you the background on where this cohort comes from, has been a very long-running longitudinal cohort, as I mentioned, based in New Zealand.

“So it followed like a thousand individuals who were born in Deniedin between 1972 and 1973.”

And the data set is really, really rich. So the researchers have been following health data on these individuals since childhood, so they're obviously in their mid-50s today. And they've undergone repeated measurements of cardiovascular health, metabolic health, cognitive testing, brain images, et cetera. And, for obvious reasons, that's a rich data set that's small and sometimes be more valuable than a cohort of a million lives for which you don't know very much, and you don't have the longitudinal tracking.

So, as we saw in the previous study, you have the Deniedin Pace Clock, which follows the longitudinal biological measurements, the blood measurements, the physiological stuff. But now the question became, "Hey, like, could we estimate a person's rate of aging just using a single measurement of something?" And, in this case, what they turned to was a neuroimaging study. Okay, so what was this? So the clock is based on a brain MRI, as opposed to blood biomarkers, and it was developed by researchers at Duke and Harvard, along with the University of Otago.

And it used data from about 875 of the 1000+ participants in the Deniedin cohort. So, at age 45, the participants underwent a T1 weighted structural MRI scan of the brain, along with a cognitive assessment. And, from those scans, the researchers extracted something on the order of 300 structural brain features, including measurements like cortical thickness, surface area, volumes of different brain regions, et cetera. They used machine learning methods to then train a model to predict the previously measured longitudinal pace of aging from those brain features.

The final model ended up relying on 99 of those 300 brain measurements. And, importantly, once it was trained, it was able to do this off just a single MRI and estimate a pace of aging. So, you do this on a training set, and then the question is, how does it predict for real? So, within the Deniedin cohort itself, the model showed a correlation of 0.6 with the longitudinal pace of aging score.

And then when the author is performed across validation, which is basically splitting the data set into a training and testing subset, the average correlation dropped to about 0.42.

So, again, what you're doing here is you don't want to use your whole cohort to build the model, because then you don't have a cohort to test it on.

“If you have a robust cohort, you have to use part of those people to build the model, and then part of those people to test the model.”

So, the final model ends up relying on 99 of these key 300 somewhat brain measurements. And, most importantly, once it's trained, it requires only a single T1 weighted MRI to establish this person's rate of aging or the pace of aging. So, how do you do this? Well, in practice, you have to take your data set, and you have to take a portion of it, and train the model on that. But you can't use your whole data set, because you have to keep part of it as an unbiased piece that you then validate the model on.

So, when they did this, they chose a portion, did the assessment, did the validation, and then they showed a correlation with their longitudinal pace of aging score that was about 0.6. And then they did something where they did a cross validation, so they just kept picking different training sets and different testing sets, and the average correlation dropped to 0.42.

At the time they did this, the people in the Dunedian cohort are like 50, 51, 52 at the most. I don't know the exact time that the study was done, but just call it a bunch of 50 year old people.

“So, you have a cohort of a little over a thousand people who are 50-ish years old, which means very few of them have died.”

None of them have got dementia, very few of them have had heart attacks, very few if any of them have become frail. And yet, you're trying to predict those things based on the subtle deteriorations in health that people will have occurred, or will have occurred in people up until the age of 50.

So the real test for the rubber hitting the road on this is how will this clock work in people who are older, which are therefore outside of your calibration window.

And to do that, they went to look at two cohorts. So one is called the Alzheimer's Disease Neuroimaging Initiative cohort ADNI.

“There were somewhere between 1300 and 1700 participants. I can't exactly tell how many they used in the study. I think they excluded 400 of them and used 1300.”

This is a North American cohort. These are patients that had either MCI, mild cognitive impairment, or even early Alzheimer's Disease. Average age was 74, but the range is 52 to 97. So this became one cohort this study. And then the other is they went to are tried and true UK biobank. And they included 42,000 people from this cohort. The age range here was 44 to 82 with a mean of 64.

And again, this excluded people with significant chronic disease. So we were not out there trying to select the second to the second. It's just give us an average 64 year old.

Okay. Again, this is a very interesting test for the clock because the clock wasn't built on people that looked like the people it's being asked to evaluate. Okay. So how did this perform? Well, in both the ADNI and the UK biobank cohorts, faster aging according to the model was associated with more rapid hippocampal atrophy, which is a well known structural change seen in Alzheimer's Disease. In the UK biobank cohort, faster than Eden pack NI scores. So that just again is the rate of aging based on this brain MRI. We're associated with greater frailty about 14% higher risk of developing major chronic diseases like diabetes and heart disease and a roughly 30% higher risk of death during follow up.

And Alzheimer's disease neuro imaging initiative data set, faster scores were also linked to substantially higher risks of developing dementia, and not surprisingly, because those people are coming in with MCI many of them. The researchers also compared their model to existing MRI based brain aging predictors and found that the the need and pack NI performed at least as well and in some cases better at predicting mortality. So I think when you take these things together, I found that to be very impressive given that the the need and pack NI was based on the brains of people who were 50 didn't have dementia and was still able to extract the changes that are occurring in people between the ages of say 20 and 50.

“And predicting things like frailty dementia and CVD. Again, maybe I'm the only one that finds that really exciting, but but I think that is awfully cool. Okay, so what is this add to the discussion.”

Well, I think one interesting take away from this study is even though the model was based on a single brain scan. It was designed to approximate the pace of aging, which is basically a whole body measurement, so maybe the old added that what's good for the heart is good for the brain and vice versa is like maybe the brain is kind of a canary in the coal mine for health, right, maybe if all of the subtle changes we can see in a brain are telling us about frailty, I mean again, that's very counterintuitive, but that's exactly what what they found here.

So there's something about the brain that is so interesting that if we can measure small things and it might be telling us something that's going on. And of course, it raises an interesting question, which is perhaps different organs can carry, I don't know, distinct, but yet overlapping signals of systemic aging. And maybe ultimately we want somebody to be able to do this for the kidneys and the liver and the brain and we'd like to see to certain disease states become more predicted by the rate of aging and a given organ to certain organs become more predictive of other things.

Unfortunately, at the level of everybody needing to know this, it's not going to happen. And I think the blood-based tests are far more likely to catch on both individuals and frankly even for clinical trials. Second, I think that while the model shows statistically significant associations with health outcomes, the predictive strength was not huge. It was moderate. By the way, we are going to include the figures to the papers in the in the show notes that you can actually look at the actual effect sizes.

It goes back to the same issue that I had with the aging clocks, which is at the population level they matter, right? So if you look at the Kaplan Meyer curves and you see that, for example, the, you know, if you take somebody who was who had a rate of pace aging that was minus one standard deviation and plus one standard deviation. So you go two standard deviations on pace aging predicted about a 1% absolute difference in mortality at seven and a half years. Okay, so we're going to, oh, I'm not going to run through all of these because it would take forever, but you can see them in the show notes, but that just gives you a sense of like that's one of the biggest markers that it found.

Hazard ratio of 32 of 1.32. So a 32% increase in mortality, but that's relative risk. So the absolute risk is about one was about 1% at seven, it's about a year. So again, at the individual level, it's not clear that that moves the needle, although with the population level with a big enough population at night. So again, where does that leave us? Well, look, faster, faster pack and I score is associated with a higher risk, but I don't think we know that slowing it means living longer at the individual level or reducing the risk of dementia.

“But look, I think it's, I think it's an exciting tool. And I think that at least my intuition is that rate of aging clocks may offer us a better insight into the question that I think we all want to know, which is what should we be doing.”

Even if I take something that I think is valuable, how should I dose it, right? Exercise would be a great example. How much exercise should I be doing? Especially if I'm a person who doesn't like to exercise that much and I just want to find that minimum effective dose.

So where does this leave us? The first study that do health trial shows that aging clocks can detect small biological changes in response to an intervention.

In this case, it was omega-3 supplementation that appeared to slightly slow several epigenetic clocks over three years.

“The second study, the Dunedin Pack and I paper showed that scientists could build new types of aging clocks, in this case using structural brain imaging,”

that appeared to capture meaningful patterns related to cognitive decline, frailty, and even overall mortality.

So taken together these studies illustrate both the promise and limitations of aging clocks. On the promise inside, these clocks may help researchers detect early biological signals in situations where traditional clinical outcomes would take decades to measure. And I think that's incredibly valuable for aging research, running a 20-year prevention trial for every possible intervention simply isn't feasible. And the biomarkers that capture aspects of the aging process could help us prioritize which ideas are worth testing more rigorously and putting more resources into.

“But at the same time, these clocks are still models and models come with limitations, in fact, to quote, "The famous physicist and it's debated if it was Fermi who said this, all models are wrong, some are useful."”

But the question is how useful are these models? Well, different clocks capture different aspects of biology, small shifts in the measurements don't necessarily translate into meaningful improvements in health outcomes.

Most importantly, we still don't know whether changing the clock actually changes what we ultimately care about things like disease risk, disability, or lifespan.

This uncertainty matters a lot when we move from research into consumer health. Right now, aging clocks are increasingly being marketed as a tool for individuals, something you can order online, track over time, and use to evaluate your own lifestyle changes. Some of these companies even promise to improve your biological age with supplements that they will conveniently sell to you as well. But based on the current evidence, it's not clear that these numbers actually give consumers any actionable information.

If your aging clock changes by a few months, what should you do differently, should you change your diet, your exercise program, your medications, take more of the supplements?

And in many cases, we already have much more reliable metrics that tell us about health and longevity risk.

“Things like blood pressure, glucose, lipids, whether you're smoking or not, all the various metrics we have around physical fitness and body composition.”

These are not particularly glamorous biomarkers, but they have something that aging clocks don't have yet. Decades of evidence linking them directly to real clinical outcomes.

In fact, if you take us back for a moment, we've already solved a large part of the problem that aging clocks are trying to address.

There's a particular industry out there that is so good at doing this that their formulas are proprietary. Life insurance companies have been predicting mortality risk for decades using actuarial models based on these various factors. In fact, I recently reached out to a senior member of a life insurance company, and this person shared with me that if they ever see a deviation in expected premium payouts that exceed one percent, it would be considered the most unusual event they could imagine.

So that means that at the population level, they have to be able to predict mortality at a degree of accuracy that exceeds anything we can imagine.

And I further asked if they were using any of the commercially available or research grade biological clocks, and the answer was no, so I think that tells you something that these companies are still doing their jobs surprisingly well, and they don't use any of the aging or biological clocks. They rely instead on the data that we understand.

“So at least for me, I think about this as if someone were to offer a biological age score, it's worth asking them what that number is actually telling them.”

Is it telling them something new or at best, is it just repackaging information you already have or understand? So again, this is not to say that these clocks provide no value. They are fascinating scientifically, and they may become more valuable as research tools over time. They may evolve into clinically meaningful biomarkers over time. But right now, I would say they would be best viewed as experimental tools for studying aging, at least at a broad enough population level, but not as definitive health metrics for individual decision making.

If you're interested in your own longevity, take away is quite simple, instead of obsessing over whether your biological age is 42 or 45, it's probably much more productive to focus on the things you already know are going to matter. Right, staying active, eating a balanced diet, getting appropriate sleep, maintaining and measuring clinically validated biomarkers. Now again, they might not sound as flashy as biological age and the scores that are attached to them, but they remain some of the most powerful tools that we have for improving both lifespan and health span.

And as aging research continues to evolve, perhaps one day these biomarkers that we get out of these clocks will help guide those efforts even further. But I think for now, it's safe to say that the fundamentals still matter most.

“Thank you for listening to this week's episode of The Drive, head over to Peter AtiaMD.com/shownotes if you want to dig deeper into this episode. You can also find me on YouTube, Instagram and Twitter, all with the handle Peter AtiaMD.”

You can also leave us review on Apple Podcasts or whatever podcast player you use. This podcast is for general informational purposes only, and does not constitute the practice of medicine, nursing, or other professional health care services, including the giving of medical advice. No Dr. Patient Relationship is formed. The use of this information and the materials linked to this podcast is at the user's own risk. The content on this podcast is not intended to be a substitute for professional medical advice, diagnosis or treatment.

Users should not disregard or delay an obtaining medical advice from any medical condition they have, and they should seek the assistance of their health care professionals for any such conditions. Finally, I take all conflicts of interest very seriously. All of my disclosures in the company's I invest in or advise please visit [email protected]/about where I keep an up-to-date and active list of all disclosures.