Deep divefrom Episode 6Longevity12 October 202626 min

Not Just Older, Noisier — DNA Methylation and Aging

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

The paper: DNA methylation variability provides a complementary epigenetic signature of aging heterogeneity (Journals of Gerontology, Series A)

MP3

Chapters

Transcript

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

Read the whole episode

Cold open

Host 1You heard the trailer — this is the whole story. Welcome to the Hastrup Review of Science — where artificial intelligence reads real science, so you don't have to pretend you understood the paper. One paper today: blood from about thirteen hundred Canadians, and the idea that how scattered your DNA chemistry is might say something about how you are aging.

Host 2The paper is called DNA methylation variability provides a complementary epigenetic signature of aging heterogeneity. It is by Olga Vishnyakova, Lloyd Elliott, Angela Brooks-Wilson and colleagues at Simon Fraser University and BC Cancer in Vancouver, with the frailty researcher Kenneth Rockwood of Dalhousie University, and it is now published in The Journals of Gerontology, Series A.

Host 1And there is a wrinkle about which version we read.

Host 2The published title also names the Baltimore Longitudinal Study of Aging, which suggests the journal version added an American cohort. The full text we could obtain is the preprint from August twenty twenty-five, which uses only the Canadian data. Every number today comes from that preprint, so anything about Baltimore is, for us, unreported.

Host 1The claim, in a sentence or two?

Host 2Most aging research asks whether a chemical mark on DNA goes up or down as health declines; this team also asked whether it becomes more scattered between people. The two questions point to largely different places in the genome, and a predictor built from both tracked death and lung disease in a small test group somewhat better than well-known epigenetic clocks.

Host 1Somewhat better is doing some work in that sentence. Let's earn it, starting with the basics: what is DNA methylation?

Why this field exists

Host 2As established background, nearly every cell in your body carries the same DNA, yet a liver cell and a white blood cell behave completely differently. Part of the reason is chemical tags on top of the DNA that help decide which genes are read. The best studied is methylation: a small methyl group attached to the letter C, usually where a C sits right before a G, a spot called a CpG site.

Host 1And the word epigenetics?

Host 2The developmental biologist Conrad Waddington coined it in the nineteen-forties for how genes and environment together shape a developing organism. Today it mostly means changes in how genes are used that leave the DNA sequence itself untouched. Methylation is the workhorse because it is chemically stable and can be measured from an ordinary blood sample.

Host 1Measured how, and how much of it?

Host 2This study used a standard commercial chip, the Illumina EPIC array, which reads methylation at roughly eight hundred and sixty thousand CpG sites per person. For each site you get a number between zero and one: the fraction of DNA copies in the sample that carry the tag at that spot.

Host 1A fraction of copies. So it is never simply on or off.

Host 2Right. A blood sample contains millions of cells, so each value is an average over all of them, and that detail matters later.

Host 1Why did aging researchers get interested?

Host 2Because methylation changes with age in a remarkably regular way. As background, between about twenty eleven and twenty thirteen several groups, most famously Steve Horvath and Gregory Hannum's team, showed that a weighted set of CpG sites could estimate a person's age to within a few years. These became known as epigenetic clocks.

What was known

Host 1Estimating someone's age seems of limited use. You could just ask them.

Host 2Which is exactly where the field went next. The interesting part was the error: people whose clock ran ahead of their real age tended to be less healthy, so later clocks were trained on health rather than age. PhenoAge, from Morgan Levine and colleagues in twenty eighteen, was trained on clinical blood markers, and GrimAge, from Ake Lu, Horvath and colleagues in twenty nineteen, on time to death using stand-ins for smoking and blood proteins.

Host 1And those are the benchmarks in this paper.

Host 2Yes, PhenoAge and two versions of GrimAge. The paper notes these kinds of clocks have been linked to mortality, heart disease, type two diabetes, cancer and Alzheimer's disease, and that some shift with interventions such as long-term calorie restriction in the CALERIE trial.

Host 1So where is the gap?

Host 2The authors' point is that all these clocks, however they are built, work on average methylation levels. At each site the question is whether sicker people carry more or less of the tag. Almost nobody asks whether sicker people are more spread out.

The open question

Host 1Spread out how? Give me a picture.

Host 2Take one CpG site across a thousand healthy people: almost everyone sits near, say, eighty percent methylated, in a tight cluster. Now take a thousand frail people: the average might still be eighty percent, but some are at sixty and some at ninety-five. A test of averages sees nothing; a test of variance sees a large difference.

Host 1Is there any reason to think that spread matters biologically?

Host 2There is a line of evidence, cited in the paper. In twenty ten Andrew Feinberg and Rafael Irizarry argued that random epigenetic variation is itself a driver of disease, and in twenty eleven Kasper Hansen and colleagues found that cancers show far more variable methylation than normal tissue. Back in two thousand five, Mario Fraga's team showed that identical twins drift apart epigenetically as they age.

Host 1Identical genes, diverging marks.

Host 2Exactly, and the authors add two more specific clues. A twenty twelve comparison of newborns and centenarians found that sites differing in average and sites differing in variability barely overlapped. And this team's own earlier work suggested that variance differences between the least and most healthy people flag traits under tight bodily regulation.

Host 1So the open question is?

Host 2Whether variability in methylation tracks how well people are aging, not just how old they are, and whether it adds anything averages miss. Concretely: where in the genome does it show up, does it overlap with the average-level signal, and can it improve a predictor of who gets ill or dies.

The cohort

Host 1Who were the people?

Host 2The Canadian Longitudinal Study on Aging, a national study of about fifty-one thousand Canadians aged forty-five to eighty-five, recruited across the ten provinces. A subgroup, the Comprehensive cohort, had detailed physical exams and gave blood, with baseline visits between twenty twelve and twenty fifteen and follow-up averaging six years.

Host 1And all of them had methylation measured?

Host 2No, a small slice. Of about twenty-seven thousand who gave blood, one thousand four hundred and seventy-eight were profiled, one thousand four hundred and forty-five passed quality control, and the main analyses used the one thousand two hundred and eighty-four unrelated participants of European ancestry.

Host 1Why only European ancestry?

Host 2That group was eighty-nine percent of the sample, and restricting to one ancestry keeps genetic background from leaking into the methylation signal. The cost is that the results may not carry over to anyone else, which the authors acknowledge.

Host 1Who was left out of the cohort from the start?

Host 2By design: residents of the territories and some remote regions, people living on First Nations reserves or in nursing homes, full-time military, people who could not respond in English or French, and anyone with significant cognitive impairment at recruitment. That last exclusion matters for the cognitive results.

Host 1Average age?

Host 2Sixty-four, give or take ten years, and fifty-one percent women.

Measuring how people age

Host 1Now the outcome. How do you turn aging well into a number?

Host 2Three ways, each scaled from zero for healthiest to one for least healthy. The first is a Frailty Index: take fifty-one possible health deficits, such as chronic conditions, count how many a person has, and divide by fifty-one.

Host 1That sounds almost too simple.

Host 2As background, it is deliberately simple. The deficit-counting approach was developed by Kenneth Rockwood, a co-author here, with Arnold Mitnitski around two thousand one, and its central finding since is that the proportion of things wrong with you predicts death and disability better than age does.

Host 1And the other two?

Host 2Cognitive function, averaged across tests of memory, executive function and psychomotor speed. And physical function, averaged across five physical tests, with each test's scores ranked and normalized before combining.

Host 1How closely do these follow age?

Host 2Fairly closely: correlations with age of roughly zero point four to zero point five. Each score carries a lot of age, which is why every model adjusts for it.

Two questions at every site

Host 1Now the analysis itself. Walk me through it.

Host 2First they split people at random: eighty-five percent, one thousand and ninety-one people, for discovery and training, and fifteen percent, one hundred and ninety-three people, held back for testing. Then, for each of about seven hundred and eighty-three thousand sites that passed filters, they asked two questions.

Host 1Question one.

Host 2Does the average methylation at this site shift with the health score? That is a standard epigenome-wide association study, run with a widely used tool called limma, adjusting for age, sex and three genetic ancestry components.

Host 1Question two.

Host 2Does the spread at this site change with the health score? For that they used double generalized linear models, which fit two equations at once: one for the average and one for the logarithm of the variance, each including the health score and the same adjustments.

Host 1Why is that a step forward?

Host 2Earlier variance methods, such as one called DiffVar, compared discrete groups, say old against young. Here the health score is continuous, so nobody has to chop people into healthy and frail bins and throw information away.

Host 1Let me pin something down. Variability here means between people, not one person's methylation wobbling over time.

Host 2Correct, and it is easy to mix up. There is one blood sample per person; the finding is that among people with worse health scores, values at certain sites are more scattered across individuals.

Host 1And then you go from sites to regions.

Host 2Neighboring CpG sites tend to move together, so single sites are noisy. A tool called DMRcate smooths the statistics over about a thousand letters of DNA and merges nearby significant sites, giving differentially methylated regions for averages and differentially variable regions for spread.

Host 1Any housekeeping worth mentioning?

Host 2Two pieces. They removed probes sitting near common genetic variants, which can fake methylation differences, and the data were adjusted for the estimated mix of immune cell types in each sample, since a frail person may simply have a different blend of white cells.

Host 1Did they test anything else?

Host 2They looked for sweet spots: sites where both too much and too little methylation go with worse health, using a regression with a single bend. Their earlier paper had found optimal ranges like that for physiological measures, so it was a fair question to ask.

Building the biomarker

Host 1And the predictor.

Host 2Built in two steps, on the training group only. Step one: for each health score, a penalized regression called elastic net selects a subset of the significant sites, average-type and variance-type together, and learns to predict that score from methylation.

Host 1So you end up with three methylation-based estimates: frailty, cognition and physical function.

Host 2Yes, each rescaled to look like years of age. Step two borrows GrimAge's recipe: a survival model weights those three estimates by how well they predict death, the weighted sum becomes an epigenetic age, and what remains after removing chronological age is the final score, called epigenetic age deviation.

Host 1A positive deviation means your blood looks older than you are.

Host 2That is the idea. For comparison they built versions using only average-type sites and only variance-type sites, a simple methylation risk score, and a measure of methylation disorder called Shannon entropy, alongside PhenoAge, both GrimAges and a null model that knows only age and sex.

Host 1All judged on the held-out people.

Host 2On the held-out group, using survival models adjusted for age and sex. The main yardstick is the concordance index: for any two people, how often the model correctly ranks who dies first, where zero point five is a coin toss and one is perfect.

Results: the averages

Host 1Results. First, the averages.

Host 2Two thousand seven hundred and forty-two single sites passed the false discovery rate threshold, and fifty-four passed the much stricter genome-wide threshold. Grouped together, they formed four hundred and forty-eight differentially methylated regions.

Host 1Anything recognizable among the top hits?

Host 2One that anyone in the field will recognize: a site in a gene called G P R fifteen, where frailer people had markedly lower methylation. As background, that exact site is one of the best-known smoking markers in blood.

Host 1So part of this frailty signal could be smoking.

Host 2Possibly, and we will come back to it. Other top hits sit in immune genes such as interleukin nineteen, and the strongest region for both frailty and physical function overlaps a gene called S L C one A five, which transports amino acids into cells and has been tied to immune regulation.

Host 1And the big picture across all four hundred regions?

Host 2When they asked which biological functions were overrepresented, sixty-seven gene ontology terms came out, led by immune system process, inflammatory response and regulation of the immune system. The top pathway in a second database, KEGG, was Epstein-Barr virus infection, followed by blood cell development.

Host 1Epstein-Barr, the glandular fever virus.

Host 2Yes, which most adults carry for life. The authors read the overall pattern as immune function being a key driver of differences in aging, which fits a long-standing idea in the field sometimes called inflammaging.

Host 1Did the three health scores behave alike?

Host 2No. Frailty dominated, with two hundred and forty-four genes unique to it and sixty-seven shared with physical function, while cognition produced only two significant regions.

Results: the spread

Host 1Now the new part: the spread.

Host 2Six hundred and ninety-two single sites passed the strict genome-wide threshold for variability, against fifty-four for averages, and four hundred and sixty-nine of those were more variable in less healthy people. At the looser false discovery threshold there were five thousand seven hundred and fifty-one.

Host 1So variability produced more hits.

Host 2Many more at the strict level, and they merged into four hundred and eighty-eight variable regions touching four hundred and forty-three genes. Unlike the average-level results, these were widely shared across health scores: sixty-six genes came up for all three.

Host 1Any standout gene?

Host 2One called PSENEN appears for frailty, cognition and physical function alike. It encodes part of the gamma-secretase enzyme complex, which is involved in Notch signaling, a pathway that helps regulate immune responses.

Host 1And the pathway analysis for the variable regions?

Host 2Nothing survived correction. The variable regions did not group into any recognizable biological function, which is worth saying plainly.

Host 1What about cognition, which had only two average-level regions?

Host 2It had plenty of variable regions, some with very strong statistics. The authors suggest variability may catch early changes before averages move, while noting that people with significant cognitive impairment were excluded at recruitment, so this concerns differences within the normal range.

Results: two different maps

Host 1The central claim: do the two maps overlap?

Host 2Barely. Only fifty-two individual sites were significant in both, and at the gene level there were four hundred and thirty-one genes near average-type regions only, four hundred and twenty-two near variable regions only, and twenty-one in both.

Host 1Do they sit in different kinds of DNA?

Host 2Average-type sites were relatively enriched in so-called open sea, stretches away from CpG-dense islands, while variable sites were enriched inside CpG islands, which often sit at gene control regions. Variable regions were also smaller: about three CpGs and three hundred and fifty letters on average, versus about five CpGs and five hundred letters for average-type regions.

Host 1And the sweet spots?

Host 2None. No bend-shaped relationship survived correction, so in these data methylation and health relate along a straight line rather than around an optimal middle. That contradicts the team's own earlier finding for physiological measures, and they say so.

Results: the biomarker

Host 1The predictor. How did it do?

Host 2The final model used five hundred and ninety-four sites. Only eight of them are shared with GrimAge and PhenoAge, and only two genes appear in all three clocks.

Host 1And against death?

Host 2In the test group, one hundred and eighty-seven people with complete data, fourteen died over about six years. Each one-unit rise in the score, which is scaled like a year of age, went with roughly a twelve percent higher hazard of death, with a confidence interval from about one and a half to about twenty-four percent.

Host 1Concordance?

Host 2Zero point eight seven seven for the combined model. Average-only sites gave zero point eight seven four, variance-only zero point eight five five, GrimAge version two zero point eight two nine, and PhenoAge zero point eight zero one.

Host 1And the null model, the one that knows only age and sex?

Host 2Zero point seven nine six. That number is the key to reading all the others.

Host 1Why?

Host 2Because most of the predictive skill comes from age and sex alone; the combined model adds about eight hundredths on top. And the gap between the combined model and the average-only model is three thousandths, with standard errors of nearly four hundredths.

Host 1So the claim that combining the two beats averages alone...

Host 2Is not demonstrated by these numbers. The direction is consistent with it, but the difference is far smaller than the uncertainty.

Host 1And GrimAge and PhenoAge?

Host 2Neither was significantly associated with death in this test group. Their hazard ratios pointed the expected way, but the intervals included no effect.

Host 1The second outcome was lung disease.

Host 2Chronic obstructive pulmonary disease, C-O-P-D. Among one hundred and fifty-seven people with data, nineteen developed it within six years, and the hazard was about sixteen percent higher per unit of the score, which held after correcting for testing several diseases. There was no association with cancer or diabetes.

Host 1What did the score itself correlate with?

Host 2The strongest links were with smoking and with two blood markers of inflammation, C-reactive protein and interleukin six.

Host 1Smoking causes most C-O-P-D, and the score tracks smoking.

Host 2Which is the obvious question, and the preprint does not report a C-O-P-D model adjusted for smoking. So we cannot tell how much of the lung signal is the score re-detecting smoking history.

Host 1Last result: genetics.

Host 2They ran a genome-wide association study on the score and estimated that common genetic variants explain about forty-one percent of its variance. The standard error is thirty-seven percent, though, so the honest reading is anywhere from almost nothing to most of it, as the authors acknowledge.

A word from our sponsor

Host 1Before the big picture, a quick word from our sponsor.

Host 1Tired of stubble on your legs two days after shaving? Then you are in luck. The Hastrup Review of Science is brought to you by epilators.

Host 2A row of tiny rotating tweezers that remove each hair from the root, so the result lasts for weeks rather than days.

Host 1I use mine every Sunday evening. The first few minutes build character, and after that it is honestly quite meditative.

Host 2Available corded or cordless, some even for use in the shower, widely in stock and reasonably priced.

Host 1Smooth for weeks. My partner noticed the difference right away.

Host 2Epilators. Our referral link is in the show notes.

Why it matters

Host 1So what is the principle here?

Host 2That an average can hide part of the story. When a biological system starts losing control, an early sign may be a loss of consistency rather than a shift in the typical value. This paper shows that in blood methylation, spread and average point at largely different parts of the genome.

Host 1Is that idea new?

Host 2As background, no, and that counts in its favor. Physiologists and ecologists have long treated rising variance as a warning sign of a system losing stability, and cancer epigenetics found methylation variability years ago; what is new is applying it to ordinary aging with continuous health scores in a national cohort.

Host 1Who does this change things for?

Host 2Mainly people building aging biomarkers. The message is to screen for variance as well as averages, because otherwise hundreds of health-linked regions are simply never looked at.

Host 1And for a listener who has paid for an epigenetic age test?

Host 2Nothing changes yet. This is a research model developed and tested within one cohort, it is not a product, and its mortality result rests on fourteen deaths.

Host 1What does it not mean?

Host 2It does not mean a noisier methylome causes aging, since this is one blood draw per person and the results are associations. It does not mean this score beats GrimAge in general, only that GrimAge underperformed in one small test group. And it does not show that the variance sites add predictive value beyond the average sites; that comparison was essentially a tie.

Honest caveats

Host 1Caveats. You have started already; let's be systematic.

Host 2First, size. Fourteen deaths and nineteen new C-O-P-D cases are very few events, and the authors themselves flag the small test set and the risk of overfitting.

Host 1Second.

Host 2The comparison with established clocks is not like for like: the new score was trained on this cohort's data and processing, while GrimAge and PhenoAge were built elsewhere. The authors suggest the established clocks suffered from calibration and preprocessing differences, and say external validation in independent data is crucial.

Host 1Is there a specific way processing could hurt those clocks?

Host 2As background rather than something the paper tested: these data were adjusted for blood cell composition before analysis, and some established clocks partly draw on cell-composition signals. Adjusting first could blunt them, which would be a plausible, unconfirmed reason for their weak showing here.

Host 1Third.

Host 2Smoking. A top average-level hit is a classic smoking marker, the score correlates with smoking, and C-O-P-D is mainly a smoking disease, yet the preprint shows no outcome models adjusted for smoking. Part of the signal may be a sophisticated smoking detector.

Host 1Fourth.

Host 2Variance is harder to measure cleanly than an average. Extra spread at a site can come from a few outlier samples, technical artifacts, leftover cell-type differences, or variance naturally changing with the average level. Beyond its general quality control and the choice of a variance-stabilizing scale, the preprint does not describe specific checks for these.

Host 1Fifth.

Host 2Who it applies to: European ancestry only, people living in the community, no significant cognitive impairment at entry. And blood is one tissue; brain or muscle may tell a different story.

Host 1Any internal inconsistencies?

Host 2A few small ones in the preprint. The sample is given both as one thousand four hundred and forty-five and one thousand two hundred and eighty-four, the test set as one hundred and ninety-three and one hundred and eighty-seven, and the discussion places average-type regions in CpG islands when the results place the variable ones there.

Host 1And the Baltimore question.

Host 2If the journal version added the Baltimore study as a replication cohort, that would address the biggest weakness, external validation. We have not seen those results and cannot tell you whether anything replicated.

What to watch

Host 1So what should we watch for next?

Host 2First, the Baltimore results in the published paper: whether the combined score holds up in an independent cohort with different processing. Second, a head-to-head of combined, average-only and variance-only models in a group with hundreds of deaths, which is what it takes to separate a few thousandths from noise.

Host 1And beyond this team?

Host 2Whether variance-type sites get folded into the next generation of clocks, and whether they move with interventions. The Canadian study keeps following its participants, and repeat blood samples could eventually show whether one person's sites become more erratic over time, which the current single-sample design cannot test.

Host 1Bottom line?

Host 2A careful, modest study with a useful idea: look at the spread, not only the average. The finding that the two point to different genes looks solid; the claim that the combined score beats established clocks rests on fourteen deaths and needs outside confirmation.

Sign-off

Host 1That is the deep dive: DNA methylation variability provides a complementary epigenetic signature of aging heterogeneity, by Vishnyakova, Elliott, Brooks-Wilson and colleagues, in The Journals of Gerontology. The link is in the show notes.

Host 2Every Friday this same feed carries the roundup episode: five new papers as trailers, and you pick which become deep dives. This has been the Hastrup Review of Science.