Most anti-aging news arrives as a promise. This one arrived as a map.
On 26 June 2026, Nature Aging published a study that did not give a single compound to a single living thing. Instead, a team led by the network scientist Albert-László Barabási at Northeastern University, working with clinicians at Brigham and Women’s Hospital and Harvard Medical School, took every drug that already exists — 6,442 approved or experimental compounds — and asked a narrower question than usual: which of them are even positioned to interfere with aging?[1]
Within days the result had been compressed into a headline about aspirin and a nasal decongestant. That is understandable, because one of the named candidates is oxymetazoline, the active ingredient in over-the-counter sprays like Afrin. It is also close to the opposite of what the paper is for. The useful thing here is not a shopping list. It is a method for separating a drug that merely touches aging biology from one that pushes it in the right direction — and, more unusually, a published estimate of how often that method gets it wrong.
Why aging is a hard target to aim at
Aging is not a disease with a receptor. Since 2013, researchers have organised it into “hallmarks” — discrete biological failure modes such as genomic instability, telomere attrition, mitochondrial dysfunction and cellular senescence. The canonical list was expanded to twelve in 2023.[2] Almost every longevity drug in clinical trials today aims at one of them, sometimes two.
That is the structural problem the new study starts from. The authors pulled 2,358 aging-associated genes from OpenGenes, a curated database that grades each gene-to-aging link from level 1 (strongest) down to level 5.[3] Only 26 of those genes carry the top grade — a sobering ratio that tells you how much of aging genetics is still suggestive rather than settled.
Some 1,250 of them could be assigned to at least one of eleven hallmark categories: 860 to exactly one, 390 to several. TP53, the tumour-suppressor gene, shows up in seven. And when the hallmarks were compared pairwise, 47 of 55 possible pairs shared significantly more genes than chance allows.
That is the quiet finding underneath the flashy one. The hallmarks of aging are not independent compartments; they are overlapping neighbourhoods in one network. A drug aimed at senescence is, whether anyone intended it or not, also tugging on inflammation and nutrient sensing.
Step one: does the drug even reach the neighbourhood?
To exploit that, the team mapped the genes onto the human interactome — the full graph of which proteins physically bind which — and found that the genes of a given hallmark cluster into a connected subgraph they call a hallmark module. That was not guaranteed, and it is what makes the rest possible: if a hallmark occupies a definable region of the network, you can measure how far any drug’s targets sit from it.
Doing this for all 6,442 compounds produced 370 drugs sitting statistically closer to at least one hallmark module than chance predicts. Notably, 83 of them are what the authors call network drugs: they do not hit a single aging-associated gene directly. They act on a neighbour, and the effect propagates. Any screen looking only at direct drug–target overlap would have missed all 83.
Step two: which direction is it pushing?
Proximity tells you a drug is in the room. It says nothing about whether it helps. This is where earlier repurposing efforts stalled, and where the paper makes its actual contribution: a metric the authors name pAGE.
The logic is straightforward. Aging leaves a fingerprint in gene expression — certain genes reliably turn up, others turn down, as tissue gets older. Drugs leave fingerprints too, catalogued for over a million perturbations in the Connectivity Map database built at the Broad Institute.[4] Lay one over the other. If the drug’s changes run opposite to the aging changes, pAGE is positive. If they run with them, pAGE is negative — the drug is reinforcing the pattern, not opposing it. Proximity plus pAGE together form the pipeline they call SHARP.
The number that did not make the headlines
A prediction engine is only worth the error rate it publishes, and to their credit the authors published theirs.
They first tested SHARP against drugs already known to extend mouse lifespan through the US National Institute on Aging’s Interventions Testing Program, the closest thing the field has to a referee.[5] Of eleven ITP-confirmed lifespan extenders, eight had expression data — and SHARP flagged all eight, for 100 percent sensitivity, though with eight compounds the confidence interval runs from 63 to 100 percent. Against seventeen compounds already in human longevity trials, sensitivity was 88.9 percent. Against ten compounds tested independently after the analysis was finished — a genuinely blind test, and the strongest evidence in the paper — it again caught every one that worked.
Then comes the part the coverage skipped. The authors also ran the seventeen ITP drugs that had failed to extend mouse lifespan. Seven had expression data, and SHARP wrongly flagged three as candidates: a false-positive rate of 42.8 percent, with a confidence interval stretching from roughly 10 to 82 percent.
In plain terms: this tool is good at not missing real hits, and mediocre at rejecting duds. Something close to half of any candidate list it produces may be noise, and with a sample of seven we cannot yet say whether it is closer to a tenth or four-fifths. That is not a criticism of the paper, which states this openly. It is a criticism of every write-up that quoted the candidate names and left out the error bar.
What actually came out of the machine
Of the 370 proximity hits, only 60 had the expression data needed to compute pAGE. Twenty-one scored positive, making them the paper’s prime candidates; 23 scored negative; 16 were inconsistent. The remaining 310 have no expression profile at all, and the authors extrapolate that perhaps 108 might prove beneficial — labelling that clearly as an extrapolation.
Individual results are more interesting than the aggregate. Aspirin’s targets sit significantly close to six hallmarks and dasatinib’s to five — dasatinib being half of the dasatinib-plus-quercetin pairing familiar from senolytics research. Meanwhile sirolimus, better known as rapamycin and the best-evidenced lifespan extender in mammals, reached significance for exactly one hallmark: altered intercellular communication.
That asymmetry is worth sitting with. The drug with the strongest real-world longevity data is the narrowest on this map, while a painkiller is the broadest. Either network breadth is not what determines whether a drug extends life, or the map is missing whatever makes rapamycin work. Both readings are uncomfortable, and the paper does not resolve them.
The worked example, and why it is the best part
Oxymetazoline shows what the approach is actually good for. It is an adrenergic agonist targeting ADRA1A, ADRA1B and ADRA1D plus three serotonin receptors — and ADRA1A is itself a top-confidence aging gene, with constitutively active ADRA1A already linked to lifespan extension in mice.
When the researchers traced how oxymetazoline’s perturbation actually travels through the network, it did not take the obvious short route. It propagated through ACKR3 — a protein that is not an aging gene at all — which then reaches NFKB1, TP53 and AKT1, counteracting age-related changes in a cluster of inflammation genes including CCL5, RELA and NFKBIA, for a pAGE of 0.46.
That is not a health claim. It is a falsifiable mechanistic hypothesis, specific enough that a lab can go and prove it wrong. Given how much longevity messaging consists of unfalsifiable gestures at “cellular health,” that is a meaningful shift in kind.
The list nobody put in the headline
The 23 negative-pAGE compounds are, journalistically, the more arresting half. They include the alpha-blockers prazosin, doxazosin and alfuzosin, prescribed to millions for blood pressure and prostate symptoms; the anti-inflammatory ketorolac; the antibiotic minocycline; benzatropine, used in Parkinson’s; pimecrolimus, an eczema cream; and navitoclax, an experimental senolytic that appears on the age-accelerating side for stem-cell exhaustion despite being developed to clear senescent cells.
Three cautions, in order of importance.
First, and non-negotiably: these are transcriptional signatures measured in cultured cell lines — much of the Connectivity Map runs on MCF7 breast cancer cells — not health outcomes in human beings. There is no evidence here that anyone’s blood pressure medication is aging them. Stopping a prescribed drug carries certain, immediate, well-documented risk; this list carries a hypothetical one. The asymmetry is not close.
Second, the 42.8 percent false-positive rate presumably cuts both ways. If nearly half the pro-longevity list may be spurious, the same should be assumed of the warning list.
Third, an internal oddity the paper does not dwell on but a careful reader should. For the intercellular-communication hallmark, oxymetazoline scores positive — and so does terazosin, which acts on the same alpha-1 receptors in the opposite pharmacological direction. Meanwhile prazosin and doxazosin, terazosin’s close chemical relatives, score negative. Whatever pAGE is capturing, it is a downstream transcriptional consequence, not a clean readout of receptor logic. That is a reason to treat individual drug names on either list as provisional.
What would have to happen next
The authors are careful to call their output an experimentally falsifiable framework, and that is the honest description. The path from here runs cells, then mice, then people. The ITP alone takes roughly four years per compound, and it exists precisely because so many promising results evaporate at that step.
Several limitations bound the exercise. Only 1,347 of the 6,442 drugs have Connectivity Map profiles, so most of the map is dark, and pAGE varies between cell lines. Not every age-related expression change is harmful — some stress-response pathways switch on with age and are genuinely protective, a hormesis effect that would make a good drug look bad. And network proximity has no concept of dose: a compound can be geroprotective at one exposure and toxic at another, and nothing here can tell you which.
What the study delivers, then, is infrastructure rather than an answer: a way of turning two decades of aging genetics into a ranked, testable queue. That is a real contribution, and also exactly the sort of thing that gets mistranslated into a supplement stack within a week. If you want something that reflects your own biology today rather than a candidate list for 2035, the measurable end of this field — the epigenetic clocks and blood-based aging markers — is far further along than the pharmacology, and the interventions with real human outcome data remain unglamorous. Our reviews of longevity supplements worth the money and the daily habits with the strongest evidence cover ground settled enough to act on now.
Common questions
Does this study mean aspirin slows aging?
No. The study found that aspirin’s known protein targets sit unusually close to six of the eleven hallmark modules in the human protein interaction network, which makes it a plausible candidate worth testing. It did not measure aspirin’s effect on aging in any animal or person. Aspirin also carries a real bleeding risk, and a large randomised trial in healthy older adults found no mortality benefit from taking it preventively. Nothing in this paper changes that.
Which drugs did the study flag as possibly accelerating aging?
Of the 60 candidate drugs with usable gene-expression data, 23 produced expression shifts that reinforced rather than opposed the age-related pattern. Named examples include prazosin, doxazosin and alfuzosin, the anti-inflammatory ketorolac, the antibiotic minocycline, benzatropine, pimecrolimus and the experimental senolytic navitoclax. These are transcriptional signatures measured in cultured cell lines, not health outcomes in people. Nobody should stop a prescribed medication on the basis of this list.
How reliable are these predictions?
The pipeline correctly flagged every mouse lifespan-extending drug it was tested against, giving 100 percent sensitivity, and 88.9 percent sensitivity for compounds already in human longevity trials. But when the authors ran it on drugs that had failed to extend mouse lifespan, it wrongly flagged 42.8 percent of them as candidates. That false-positive rate was estimated from only seven compounds, so its confidence interval runs from about 10 to 82 percent. In practice, a meaningful share of the candidate list is likely to be wrong.