LONGEVITY LATEST ISSUE 31 · 7 OCTOBER 2026
LONGEVITY LATEST
The Evidence-Based Edge on Living Longer and Better
Issue 31 · The Machine and the Clock · 7 October 2026
WELCOME
Welcome
In a Swedish trial of 105,934 women, AI-supported breast screening found 29% more cancers while radiologists read 44% fewer images. That much could be checked within a few years. Whether it saves lives will take longer.
That gap, between what can be checked quickly and what can’t, runs through this whole issue.
Falls move to next week. September brought an AI-designed drug said to make patients “biologically younger”, and it deserved a proper look now. Inside: three AI tools graded, the $499 question of an AI-read biological age, and why AI-designed drugs have yet to show an edge where it counts, in phase 2.
SPOTLIGHT
AI is fastest where the answer can be checked
What AI does | Best human evidence so far | What’s still missing |
Spots (reads scans) | MASAI, randomised, 105,934 women: 29% more cancers found, 44% less reading, interval cancers no higher. | Proof of fewer breast cancer deaths. |
Warns (flags risk in routine tests) | Taiwan AI-ECG trial, 15,965 hospital patients: 90-day deaths 3.6% vs 4.3% when doctors received the alert. | Replication beyond one hospital network. |
Designs (finds targets and molecules) | Rentosertib: 71 patients in phase 2a, now in phase 3. Zasocitinib: FDA decision due early 2027. | An approved drug whose target AI found. Anything aimed at ageing itself. |
Measures (ageing clocks) | Clocks predict risk across large groups. | Evidence that moving your number changes your outcome. |
The rows get shakier as you read down, and that isn’t a coincidence. A mammogram reader can be scored against biopsies within months. A lung drug shows its hand within a year. A claim about slowing ageing would take decades to score, so we score it with clocks, and the clocks are themselves machine-learning models trained on other people’s data. One model marking another’s homework.
Gommers et al. MASAI. Lancet (2026) · Lin et al. AI-ECG alert trial. Nat Med (2024) · Xu et al. GENESIS-IPF. Nat Med (2025)
Checkable is the word. The closer a claim sits to a biopsy, a hospital record or a death certificate, the more weight it can bear.
THIS WEEK’S ANALYSIS
Top 3 Interventions Under the Microscope
The grades judge evidence that an AI tool improves health, not how clever the tool is.
AI-evidence grades: A = randomised trials show fewer deaths or serious events; B = randomised trials show better detection of what matters, outcome benefit unproven; C = more findings but not clearly the ones that matter, or early human data only; D = claims outrun the data.
1. AI as a second reader in breast screening | Grade B
If you’re offered AI-supported screening, say yes. Of the three tools graded here, it has the largest and best-replicated evidence.
The evidence. MASAI randomised Swedish women to standard reading by two radiologists, or to AI triage: low-risk scans got one reader, high-risk scans got two plus the software’s flags. The AI arm found 29% more cancers without more false alarms. The final analysis, in The Lancet this January, counted interval cancers, the ones that surface between screens and tend to be nastier: 82 against 93, with fewer invasive and fewer large tumours on the AI side. In Germany’s PRAIM study, 463,094 women screened in routine practice, AI support found 17.6% more cancers with no rise in recalls.
The limit. The drop in interval cancers wasn’t statistically significant; MASAI was built to show AI was no worse, and it showed that. Finding more cancers also means finding some that would never have caused harm, and nobody yet knows how many. PRAIM was observational, radiologists chose when to use the AI, and the vendor’s technology chief was an author. B, and near the top of B. The UK’s EDITH trial, around 700,000 women, should settle much of this.
Gommers et al. MASAI. Lancet (2026) · Hernström et al. MASAI. Lancet Digit Health (2025) · Eisemann et al. PRAIM. Nat Med (2025)
2. AI during colonoscopy | Grade C
Useful, but not in the way the headlines suggest. It finds more small polyps; a gain in the dangerous ones hasn’t been shown.
I expected this to be the easy B. Polyp-spotting software has more randomised trials behind it than almost any medical AI, and a meta-analysis of them found adenoma detection up 20% (RR 1.20). Look closer and the gain is almost all diminutive polyps, the tiny ones least likely to become cancer. For advanced adenomas, the ones that matter most, the rise wasn’t statistically significant (RR 1.08, 95% CI 0.95–1.22). No demonstrated gain, which isn’t the same as no gain.
Then the awkward finding. At four Polish centres, experienced endoscopists who’d used AI for three months found adenomas in 22.4% of their colonoscopies without it, down from 28.4% before. Only 19 doctors, observational, patient groups that differed: a warning light, not a verdict. But it’s the satnav problem. You arrive faster and stop learning the roads.
A randomised trial published in June, in a high-performing screening programme, found no gain at all in surveillance colonoscopy. My reading: AI lifts weaker endoscopists and barely touches the best. The question worth asking is your endoscopist’s adenoma detection rate, not whether the scope has software.
Makar et al. Gastrointest Endosc (2025) · Budzyń et al. Lancet Gastroenterol Hepatol (2025) · Codesido-Prado et al. Endoscopy (2026)
3. AI-designed drugs for age-related disease | Grade C
One to watch, not one to chase. The first drug with an AI-found target is in phase 3. Nothing is approved.
Rentosertib is the one everyone cites. Insilico’s software picked TNIK, a kinase nobody had linked to lung scarring, then generated a molecule to block it. In GENESIS-IPF, 71 people with idiopathic pulmonary fibrosis, a scarring lung disease of later life, took one of three doses or placebo for 12 weeks. On the top dose, lung capacity rose by 98 mL; on placebo it fell by 20 mL. Clean signal, tiny sample: around 18 people per arm, and the primary endpoint was safety. Diarrhoea and raised liver enzymes were the commonest side effects. A 320-patient, 52-week phase 3 dosed its first patient in China on 9 September.
This is the part that made me sit up. On 7 September, Nature Biotechnology published an analysis of 42 trial patients: six proteomic ageing clocks all read lower on the drug, by three to four years at the peak and up to six on one clock. Then I read the details. The first author is Insilico’s chief executive, the peak came at week four, and the authors concede the clocks can’t separate a slower-ageing body from a less-inflamed lung. That’s the question that matters. A falling clock can’t tell you whether the drug slowed ageing or simply eased lung disease, and the published peer-review file shows reviewers pressing on exactly that. Real human data. Not yet evidence about ageing.
Xu et al. GENESIS-IPF. Nat Med (2025) · Zhavoronkov et al. Nat Biotechnol (2026) · Insilico first patient dosed, GENESIS-IPF-3 (September 2026)
HYPE CHECK
“AI says you’re biologically 38”
The claim. A cheek swab or finger-prick, read by a machine-learning clock, reveals how fast you’re really ageing and whether your routine is working. TruAge costs $499, Elysium’s Index $299, GlycanAge $348 and TallyAge around $249.
The missing step. Clocks are models, and they don’t all measure the same thing. Yale’s TranslAGE analysis in Nature Medicine this August ran 3,128 blood samples from 51 intervention studies through 16 clocks. The newer clocks, trained on mortality or pace of ageing, responded more strongly and agreed better with one another; older clocks were less consistent, sometimes moving in opposite directions. In CARDIA, 913 adults followed for 11 years, the decades-old Framingham risk score was more strongly linked to later heart disease than any of the four epigenetic measures tested. And one researcher estimates that time of day alone can shift some results by up to five years.
In CARDIA, a free risk score was more strongly linked to heart disease than four epigenetic clocks.
None of this tests the products above directly, and the better clocks are genuinely useful research tools. But it shows what a single consumer number can’t yet carry: a decision about your own heart. For that, the measures with treatment trials behind them come first.
Sehgal et al. TranslAGE. Nat Med (2026) · Forrester et al. CARDIA. JAHA (2024) · Medscape. Epigenetic clocks: new scepticism (2025)
Our verdict. Skip it. If you’re 40 to 74 in England, the NHS Health Check is free and gives a ten-year cardiovascular risk from numbers that actually drive treatment; elsewhere, ask for blood pressure, lipids and HbA1c. A validated home blood-pressure monitor, about £30, is a better first purchase: its readings link directly to treatments with outcome trials behind them.
IN BRIEF · NEW RESEARCH
Your voice as a clock. A model published in Science Advances on 30 September extracted more than 700 features from a few minutes of speech by 2,928 Spanish speakers in five Latin American countries. People whose speech “sounded older” than their age were more likely to have cognitive impairment. Cheap and non-invasive, which is the appeal. Most people were recorded once, though, so it shows an association, not that the clock predicts decline.
A computationally designed pill nears approval. On 14 September the FDA accepted Takeda’s application for zasocitinib, a once-daily psoriasis pill, under priority review, with a decision due in early 2027. In one phase 3 trial of 693 adults, 71% hit PASI 75 against 33% on apremilast and 12% on placebo, and full head-to-head data presented at EADV on 30 September showed more than 2.5 times as many patients reaching complete clearance as on Sotyktu, the established TYK2 pill. It’s often called the first “AI-designed” drug; the scientists who designed it dispute the label. The Deep Dive explains why that matters.
THIS WEEK’S DEEP DIVE
The Phase 2 Wall
In a 2024 analysis, AI-discovered molecules passed phase 1 trials 80 to 90% of the time. In phase 2, where they first have to work, about 40% succeeded: close to historical norms, though from only ten molecules.
The companion separates the chemistry AI clearly speeds up from the biology it hasn’t yet shown it can crack, explains why comparisons with historical trial data can mislead, and gives five checks for the next “AI drug reverses ageing” headline.
Read the companion: The Phase 2 Wall →
BIOHACKING CORNER · THE SHORT VERSION
Turn up for screening. AI is already being tested inside breast screening programmes, and it only helps people who attend; do the home bowel test when it arrives too. Having a colonoscopy? Ask for your endoscopist’s adenoma detection rate and how it compares with the national standard. That number is better evidenced than any software.
My rule for chatbots and health: use them to prepare questions, never to answer them. I’ll paste a blood result, name removed, to learn what each marker means, then take the questions to someone who can examine me. If a test or supplement quotes a biological-age change, ask one question before paying: which clock, and has it been linked to health outcomes in people like you?
READER PULSE
Two yes-or-no questions: as far as you know, has an AI tool read any of your scans or tests, and would you want to be told if it had? Reply with your two answers. Last week’s DXA answers will appear in the falls issue.
CLOSING
Keep score in outcomes
Each row of the Spotlight table has a test coming. EDITH, with around 700,000 UK women, will show whether MASAI holds in another health system. Zasocitinib’s FDA decision, due early 2027, will show what AI-assisted chemistry can deliver. Rentosertib’s phase 3 will test an AI-picked target over a full year and collect blood proteins at weeks 12, 26 and 52, giving the clock question a second, larger look. I’ll grade each one as it reports.
Next week, as promised: falls, the event that turns fragile bone into a fracture, and the balance training with the strongest trial record.
Stay curious and stay healthy!
Christian Thomsen, Editor
Longevity Latest is published weekly by FrontWave Media Ltd. Educational content, not personal medical advice. Rentosertib and zasocitinib are investigational and not approved for any use. Ask a clinician which screening programmes apply to you, and don’t change treatment on the basis of a biological-age result.
Sources and reading notes
Evidence checked through 3 October 2026. One AI tool reviewed here has randomised mortality evidence (the AI-ECG alert, a single trial); the mammography and colonoscopy grades rest on detection and interval-cancer endpoints; the drug data are phase 2a. Industry involvement: Insilico authored the rentosertib clock analysis; a vendor executive co-authored PRAIM; TruDiagnostic collaborated on TranslAGE. Test prices are list prices at the time of writing.
© 2026 FrontWave Media Ltd · Longevity Latest