Deep divefrom Episode 9Longevity4 November 202623 min

Your Protein Age Moves When Your Kidneys Do

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: Intraindividual longitudinal changes of the proteome reveal dynamic markers of biological aging (bioRxiv preprint)

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. Today, one German cohort, the same few thousand people sampled up to three times over fourteen years, and the question of whether a blood-protein age can move within a single person.

Host 2The paper is Intraindividual longitudinal changes of the proteome reveal dynamic markers of biological aging, by Lermer, Schunkert, Heinig and colleagues in Munich. It is a bioRxiv preprint posted on October sixth, so it has not been peer reviewed yet. The short version is that when a person's kidney function, weight or cholesterol got worse between visits, their protein-based age estimate tended to drift upward with it.

Why this field exists

Host 1Let's start further back. Why does anyone want a biological age in the first place?

Host 2Because chronological age is a very good predictor of disease and a very blunt one. Two sixty-year-olds can have very different hearts, kidneys and life expectancies, and the number of birthdays does not tell you which is which. The idea of a biological age is a single number that captures how far along that process a body actually is.

Host 1And where did the clocks come from?

Host 2As established background, the field took off in two thousand thirteen, when Steve Horvath, and separately Hannum and colleagues, showed that DNA methylation patterns could predict age with errors of a few years. Those are the epigenetic clocks. A few years later, people started doing the same thing with proteins circulating in blood plasma.

Host 1Why proteins rather than DNA marks?

Host 2Proteins are closer to function. The plasma proteome is what the liver secretes, what leaks out of damaged tissue, what the immune system is doing and what the kidneys fail to clear. So a protein clock is, in principle, a readout of physiology rather than of a regulatory layer one step removed.

What was known before

Host 1So what was already known before this paper?

Host 2Quite a lot, from background literature. In two thousand nineteen Lehallier and colleagues measured around three thousand plasma proteins with the SomaScan platform and showed age-related waves of change. More recently, a clock called ProtAge was built in UK Biobank on the Olink platform, and organ-specific protein clocks suggested that different organs can be aging at different rates in the same person.

Host 1SomaScan, Olink. What do those have in common?

Host 2They are affinity-based. Each protein is caught by a pre-made binder, an aptamer or a pair of antibodies, so you measure a fixed menu at very high throughput. The authors point out that this limits coverage to whatever is on the menu, and that binders can be thrown off by protein variants.

Host 1And the alternative is mass spectrometry.

Host 2Yes. In liquid chromatography mass spectrometry you chop all the proteins into peptides, separate them, and identify them by their mass. It is unbiased in the sense that nobody chose the list in advance, and the peptide identification is very specific, though in plasma it tends to see fewer proteins than the big affinity panels.

Host 1And the clocks so far, were they built on people measured once or many times?

Host 2Mostly once. Cross-sectional data, meaning one sample per person. The paper notes that some longitudinal work exists, for example in the American ARIC cohort, but those studies either focused only on mortality or treated risk factors as fixed baseline predictors.

The open question

Host 1So what is the gap this paper is aiming at?

Host 2Whether biological age is a static label or a moving gauge. If you compare a diabetic person with a non-diabetic person and find the diabetic one looks older, that could be because of the diabetes, or because of genes, upbringing, income or a hundred other things that differ between those two people.

Host 1And if you follow the same person?

Host 2Then everything that stays constant about that person cancels out. If one individual's kidney function declines between two visits and their protein age rises at the same time, genes and childhood cannot explain that change. That is the question here: within one person, does the clock move when the risk factors move?

What this paper did

Host 1Walk me through what they actually did. Start with the people.

Host 2The cohort is KORA, a population study from the region around Augsburg in southern Germany. Participants were examined in three waves: S four in nineteen ninety-nine to two thousand one, F four in two thousand six to two thousand eight, and F F four in two thousand thirteen to two thousand fourteen. Roughly seven years separate each visit.

Host 1How many people, and how old?

Host 2Four thousand one hundred individuals contributed nine thousand one hundred and five plasma samples. About four thousand at the first wave, three thousand at the second and two thousand one hundred at the third, so the cohort shrinks over time. Ages ran from twenty-four to eighty-eight, with roughly equal numbers of men and women.

Host 1And the blood?

Host 2Plasma was stored at minus eighty degrees, and here is a nice detail: samples from all three waves, about nine thousand two hundred of them, were run in a single batch between December two thousand twenty-one and May two thousand twenty-two. That matters, because if each wave had been measured years apart on different instruments, a machine drift could masquerade as aging.

Host 1How much plasma does it take?

Host 2One microliter per sample. It was denatured, digested with trypsin and Lys C, and run on an Evosep One chromatography system coupled to a timsTOF Pro two mass spectrometer. Samples with more than thirty percent missing values were dropped, leaving eight hundred and forty-two proteins in the final dataset.

Host 1Eight hundred and forty-two sounds small next to the affinity panels.

Host 2It is smaller, and that is typical for mass spectrometry of undepleted plasma, where a handful of very abundant proteins like albumin dominate. The trade-off is that what you do see is measured without a pre-selected list.

Host 1Then the clock itself.

Host 2They trained a multilayer perceptron, a small neural network, to predict chronological age from the protein levels. They chose a neural network because protein levels are known to change non-linearly over the lifespan, and they compared it with an elastic net, which is the standard linear workhorse for clocks.

Host 1How did they keep it honest?

Host 2Ten-fold cross-validation, with folds defined by person. All samples from one individual land in the same fold, so the model never predicts someone's age after having seen that same person's other visits during training. Every participant's protein age therefore comes from a model that never saw them.

Host 1And age acceleration?

Host 2Predicted protein age minus actual age. If the clock says fifty-eight and you are fifty-five, your age acceleration is plus three years. Most of the paper is about what pushes that number up or down.

Between and within

Host 1Now the part you said was the point: between and within.

Host 2For each risk factor they split every person's value into two pieces. The between piece is that person's average across all their visits, and the within piece is how far each visit deviated from their own average. Both go into one linear mixed model, adjusted for age and age squared, with a random intercept per person.

Host 1In plain terms?

Host 2The between coefficient asks whether people who are generally heavier have older protein profiles than people who are generally lighter. The within coefficient asks whether, in visits where a given person is heavier than their own usual, their protein age is also higher than their own usual. The second one is the new contribution.

Host 1Which risk factors?

Host 2Body mass index, diabetes, kidney function measured as estimated glomerular filtration rate, LDL cholesterol above one hundred and thirty milligrams per deciliter, hypertension, ever smoking and physical activity. Kidney function was calculated from creatinine and cystatin C, and the authors note that neither of those was among the proteomic inputs.

Host 1Anything else in the toolbox?

Host 2Four more things: a lagged model asking whether risk factors at one visit predict age acceleration at the next. Survival analysis against deaths up to the end of two thousand twenty-one. A mediation analysis, and then two ways of looking at the proteins themselves: SHAP values to see which proteins drive the clock, and Mendelian randomization to test whether any of them might be causal.

Results

Host 1Results. How good is the clock?

Host 2Modest. The mean absolute error was six point three nine years, with a correlation of zero point eight two between predicted and real age. The authors say plainly that this is less accurate than some larger protein and methylation clocks.

Host 1Is a worse clock a problem?

Host 2Not necessarily. A clock that predicted age perfectly would leave no age acceleration to study. But it does mean a fair amount of the gap between protein age and real age is plain noise, and that noise sits inside every number that follows.

Host 1Between people first.

Host 2Diabetes had the largest effect: people with diabetes looked about two point seven years older by protein age. One standard deviation more body mass index, about four point eight units, came with roughly zero point eight years more. High LDL added about zero point six years, and ever having smoked about zero point five.

Host 1And in the other direction?

Host 2Better kidney function was the strongest protective association: one standard deviation more filtration, about eighteen and a half units, came with about one year less age acceleration. Being physically active came with about zero point eight years less. The authors note these line up with previous protein and epigenetic clocks, which is reassurance that the clock behaves sensibly.

Within one person

Host 1Now the within-person results.

Host 2The kidney signal was the largest. When a person's filtration rate was one standard deviation above their own average, their age acceleration was about one point eight years lower, and that comes with a t statistic of minus twelve point six. Within-person gains in body mass index went with about one point three years more, and crossing the high-LDL threshold with about zero point nine years more.

Host 1Diabetes?

Host 2About one year more when someone was diabetic at a visit, but with a p value of zero point zero four two, which is barely over the line. Only about four percent of samples came from people with diabetes, so there were few transitions to learn from.

Host 1What about smoking and exercise?

Host 2No significant within-person effects for the others. The authors explain why: those exposures barely changed. Ever smoking, by definition, can only switch once, and it changed for fewer than ten people in the whole cohort.

Host 1And the lagged analysis?

Host 2Risk factors at one visit predicting age acceleration seven years later. Body mass index and diabetes had the strongest positive effects, then smoking and high LDL, and hypertension also showed up. Better kidney function predicted lower future acceleration, while physical activity did not predict anything.

Mortality

Host 1Does protein age predict death?

Host 2Yes, with follow-up of up to twenty-two years. Within each decile of real age they took the top ten percent and the bottom ten percent of age acceleration and pooled them, and the survival curves separated clearly, with a log-rank p value around ten to the minus ten. The separation held at each of the three waves separately.

Host 1How much does it add on top of just knowing someone's age?

Host 2A little. The concordance index, a measure of how well a model ranks who dies first, went from zero point eight two five with age alone to zero point eight three zero with age plus protein age acceleration. Each extra year of acceleration carried about a four percent higher hazard, and that held after adjusting for kidney function.

Host 1Zero point eight two five to zero point eight three zero. That's half a point.

Host 2It is, and that is the honest size of it. Age already explains most of who dies in a population cohort, so the added information is real, statistically very clear, and small in practical terms.

A word from our sponsor

Host 1Before we get to the mediation analysis, a short break.

Host 1Has something been accelerating quietly while you were looking the other way? Is it visible from across the room?

Host 2The Hastrup Review of Science is brought to you by nose-hair trimmers. A small rotating blade on a battery, inserted carefully, and indifferent to your between-person and within-person components alike.

Host 1I used one before each of my last three visits. My trajectory was unchanged, but my residuals were much tidier.

Host 2Nose-hair trimmers are sold in most drugstores at ordinary prices. No mass spectrometer, no mixed model, no twenty-two-year follow-up. The link is in the show notes.

Mediation

Host 1Back to the paper. You mentioned mediation.

Host 2The question is whether part of a risk factor's link to death runs through protein age acceleration. Using the first wave only and fifteen-year mortality, they simulated scenarios with a method called g-computation and built confidence intervals from two thousand bootstrap samples.

Host 1What came out?

Host 2For diabetes, the total effect was about ten point three percentage points more fifteen-year mortality risk, and about two of those points ran through age acceleration, roughly a fifth. For body mass index and kidney function the share running through the clock was about a quarter. For smoking it was about four percent, and for physical activity about three percent.

Host 1Smoking barely goes through the clock?

Host 2Barely, at least this clock. Smoking's total effect was large, over five percentage points, but almost none of it was carried by the protein age signal. High LDL had no significant indirect effect at all.

Host 1Should we read that as causal?

Host 2The authors themselves say no. Mediation in observational data needs strong assumptions, including no unmeasured confounding of any of the three links, and they describe the results as consistent with a pathway through age acceleration rather than proof of one. Note also that in this analysis exposure and protein age were measured at the same visit.

The proteins

Host 1Which proteins actually drive the clock?

Host 2Ranked by SHAP values, which estimate how much each input pushes a given prediction, the top three were lumican, the acid-labile subunit of the insulin-like growth factor complex, and afamin. Lumican is an extracellular matrix protein, the acid-labile subunit is part of the growth hormone axis, and afamin is a liver-made vitamin E carrier.

Host 1Do they overlap with the other clocks?

Host 2Hardly. None of the top proteins were shared with ProtAge, and only apolipoprotein B overlapped with the Lehallier SomaScan model. The authors attribute that to platforms measuring different proteins, and argue that the pathways still converge.

Host 1Which pathways?

Host 2Among the top one hundred and fifty proteins, complement, coagulation and xenobiotic metabolism were over-represented, with proteins like vitronectin, plasminogen and apolipoprotein E. Most of those are made by the liver, so the authors read the clock partly as a readout of a changing liver secretome, alongside matrix remodeling and low-grade inflammation.

Host 1Do different risk factors age you through different proteins?

Host 2They tried to answer that with what they call residual SHAP values, an attempt to strip out the part of each protein's contribution that just reflects real age. Body mass index, hypertension and low kidney function shared a lot of their signature, with overlap scores above zero point five, and diabetes and high LDL overlapped moderately. Smoking and physical activity had signatures largely their own.

Host 1How do they strip out real age?

Host 2By assuming that chronological age is spread evenly across all eight hundred and forty-two proteins and subtracting that equal share from each. That is a convenient simplification, not something the data show, so we would treat the risk-factor signatures as exploratory.

Causal hints

Host 1And Mendelian randomization?

Host 2As background, Mendelian randomization uses genetic variants that nudge a protein's level as a natural experiment, since genes are assigned at conception and are not changed by lifestyle. Here they found genetic variants for the proteins in KORA and tested them against published genetic studies of six aging traits, including frailty, parental lifespan, healthspan and longevity.

Host 1And?

Host 2Twenty-eight of the eight hundred and forty-two proteins showed evidence consistent with a causal link to at least one trait. Fourteen of those were among the clock's top twenty percent of features, where about five or six would be expected by chance, an odds ratio of four point three. Examples include apolipoprotein E and complement component C four A.

Host 1Apolipoprotein E is not exactly a surprise.

Host 2No, and as background, the APOE region on chromosome nineteen is one of the most consistently longevity-linked regions in the human genome, and it affects many traits at once. The significance threshold for each MR test was a nominal p below zero point zero five, so individual hits should be read as leads, while the enrichment is the more robust part.

Why it matters

Host 1So what's the principle here?

Host 2That a protein-based age is not just a fixed property of a person. In the same individuals, measured seven years apart, the estimate moved in step with kidney function, weight and cholesterol, and it moved in the same direction as the between-person differences. That is what you would want from a measure meant to track change.

Host 1Who does that matter to?

Host 2Mainly to researchers who want to use aging clocks as outcomes in trials. A clock that only distinguishes people is a risk score; a clock that responds within people is a candidate for measuring whether an intervention did anything. It also matters to the mass spectrometry community, as evidence that their platform can carry a clock from a large population cohort.

Host 1And the kidney result?

Host 2The authors read it as renal and systemic aging being tightly coupled. Our reading adds a more mechanical possibility: the kidneys filter many small plasma proteins, so when filtration falls, those proteins accumulate, and a blood-protein clock may partly be measuring clearance. Beta two microglobulin, a classic filtration marker, appears among the proteins in their risk-factor signatures.

Host 1What does it not mean?

Host 2It does not show that losing weight or lowering cholesterol makes you biologically younger. Nobody was assigned to change anything; these are associations between natural changes. And it is not a test anyone should take, because an individual error of over six years swamps the one- or two-year effects being discussed.

Honest caveats

Host 1Then the caveats, all of them.

Host 2First, the authors' own: no independent cohort with repeated mass spectrometry was available, so the within-person findings have not been replicated. The population is Western European only. Plasma captures only part of aging biology.

Host 1Go on.

Host 2Second, within-person analysis removes stable confounders but not changing ones, which the authors also acknowledge. If someone develops an illness that lowers kidney function, changes their weight and shifts liver protein output all at once, the illness drives everything and the risk factor is just a fellow traveler.

Host 1Third?

Host 2Attrition. The cohort goes from about four thousand to about two thousand one hundred samples, and people who drop out or die tend to be sicker. The later visits are therefore a healthier, survivor-biased slice, which can shape within-person estimates in ways that are hard to predict.

Host 1Anything about the measurements themselves?

Host 2Several. Diabetes was defined by a single long-term blood sugar marker, glycated hemoglobin, and was rare, physical activity was a self-reported yes or no, and LDL was measured with different assays at the first wave versus the later two. The mortality gain is half a point of concordance, the mediation rests on untestable assumptions, and the risk-factor signatures rest on that equal-share simplification.

Host 1And the paper itself?

Host 2It is a preprint, and a few reported statistics in the text have obvious typesetting slips, which is normal at this stage. The code is on GitHub, while the data are available only by application to KORA because of participant consent. One senior author reports consulting and lecture fees from a long list of pharmaceutical companies.

What to watch next

Host 1What should we watch for next?

Host 2Replication of the within-person effects in another cohort with repeated mass spectrometry, ideally outside Europe. Then the real test: a trial, where a risk factor is changed on purpose, for example weight loss with a drug or blood pressure treatment, and the clock is measured before and after.

Host 1Anything on the kidney question?

Host 2A version of the clock that excludes proteins known to be cleared by filtration, or that adjusts for it, would show how much of the signal is aging and how much is plumbing. And it would be useful to see the same people run on an affinity platform too, to know whether the near-zero protein overlap is about technology or biology.

Bottom line

Host 1Bottom line?

Host 2In about four thousand Germans followed for up to fourteen years, a mass spectrometry protein age rose and fell with each person's own kidney function, weight and cholesterol, and it added a small amount to mortality prediction. It is a careful, well-designed observational study with a modest clock and no replication yet. Consistent with biological age being a moving gauge, which is exactly what the authors claim and no more.

Sign-off

Host 1That is the deep dive: Intraindividual longitudinal changes of the proteome reveal dynamic markers of biological aging, by Lermer, Schunkert, Heinig and colleagues, a preprint on bioRxiv. The link is in the show notes.

Host 1Every 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.