Sources and notes Two in three, about 2 in 100, about 3 in 1,000: EDENT1FI screening in 8 European countries (Czech Republic, Denmark, Germany, Italy, Poland, Portugal, Sweden, UK), presented at EASD 2026 by a consortium led from Helmholtz Munich. Conference data, no paper yet. 140,568 children, median age 7.6. Early-stage type 1 means two or more islet antibodies, confirmed on a second sample. 219 of 10,273 children with an affected parent or sibling (2.13%) and 408 of 127,754 without (0.32%); 408 of the 627 with known family history is 65%. The EASD release rounds these to 1 in 45 and 1 in 350; the counts work out nearer 1 in 47 and 1 in 313. Of 386 children staged, 16 already had diabetes. news-medical.net/news/202610… Over a third within 5 years: Fr1da screening of 220,476 children in Bavaria. Five-year progression from early-stage to diagnosed type 1 was 36.2%, with or without family history. Winkler et al, JAMA, June 2026. pubmed.ncbi.nlm.nih.gov/4216… Metformin first: the 2006 ADA/EASD consensus algorithm put metformin, with lifestyle change, as the first drug at diagnosis. Nathan et al, Diabetes Care, 2006. doi.org/10.2337/dc06-9912 New guidance: ADA/EASD consensus report on managing type 2 diabetes, 2026, Diabetologia, 2 October. It advocates "earlier use, potentially from diagnosis, of SGLT2 inhibitors or/and GLP-1-based therapies", and says earlier combination of the two should be considered in people with heart disease, kidney disease or heart failure. Expert consensus, not a trial. doi.org/10.1007/s00125-026-0… Dapagliflozin vs metformin, SMARTEST, results presented in a dedicated EASD session on 1 October. Open-label randomized trial run through Swedish health registers, 2,072 people with type 2 for under 4 years and no major heart or kidney disease. Main outcome: first heart attack, stroke, heart failure, new or worse eye, kidney or foot complications, or death. The Swedish diabetes society's meeting report gives a mean follow-up of 3.8 years and no statistical difference in the main outcome. No paper or event counts released yet. dagensdiabetes.se/easd-repor… pubmed.ncbi.nlm.nih.gov/4131… Orforglipron vs insulin glargine, ACHIEVE-4, Lancet, 30 September. Randomized, open-label, median 2 years, all with heart or kidney disease. Main outcome (heart death, heart attack, stroke or a hospital stay for unstable angina) 4.2% vs 5.0%, hazard ratio 0.84 (0.59 to 1.20). That meets the trial's non-inferiority bar (upper limit under 1.8, a wide margin); it doesn't show the pill is better. Lows under 54 mg/dL 6.8% vs 19.2%. Stomach side effects 62.1% vs 14.2%. Deaths 19 vs 43, outside the trial's formal tests. Funded by Lilly. doi.org/10.1016/S0140-6736(2… Retatrutide, TRIUMPH-2, Lancet, 29 September. 1,152 adults with obesity and type 2 diabetes, 80 weeks, 12 mg dose, everyone randomized counted (treatment-regimen estimand). Lilly's release gives 20.8%, an estimate as if everyone had stayed on treatment. doi.org/10.1016/S0140-6736(2… Survodutide, SYNCHRONIZE-2, NEJM, 1 October. 752 adults, 76 weeks, 6.0 mg dose, counted the same way. Stomach side effects 77.7% vs 38.6%. Boehringer Ingelheim's release gives up to 13.1%. doi.org/10.1056/NEJMoa260721… GAD-alum, DiAPREV-IT follow-up, Diabetologia, 28 September. Randomized, placebo-controlled, 50 children aged 4 to 17 with GAD antibodies plus at least one other islet antibody, two shots 30 days apart, median follow-up 12.9 years, 49 analyzed. The gene type is HLA-DR3-DQ2. The original trial missed its goal and this is a subgroup analysis: the hazard ratio favored GAD-alum in that group (0.315, p=0.018), but the 9.0 vs 3.4 year medians alone weren't statistically separated (p=0.096), and in the 22 without that gene type progression was faster on the drug (hazard ratio 3.5). Three of the seven authors work for Diamyd Medical. Published during the meeting week; not on the EASD programme. doi.org/10.1007/s00125-026-0… Mauritius: seven national surveys from 1987 to 2021, 36,744 attendances, adults aged 25 to 74, rates standardized for age and sex. Lancet Diabetes & Endocrinology, 30 September, presented at EASD 2026. Prediabetes fell from 24.9% to 15.6%, and the share of cases first found by the survey fell from 67.0% to 32.5%. The 2021 survey had fewer attendances (3,343 vs 5,833 in 2009). Surveys show the trend, not its cause. doi.org/10.1016/S2213-8587(2… Semaglutide and pancreatic cancer: new users in Danish, Swedish and Norwegian registries compared with new users of sulfonylureas, SGLT2 inhibitors or insulin. Follow-up started a year after the first prescription and averaged 1.40 to 1.85 years after that; 131 vs 123 cancers, pooled hazard ratio 0.91 (0.71 to 1.16). A safety study required by regulators and funded by Novo Nordisk, late-breaking abstract at EASD 2026, no paper yet. A 2025 US veterans study found modestly higher odds than with DPP-4 inhibitors (0.24% vs 0.17%), so the question isn't fully closed. news-medical.net/news/202609… doi.org/10.1016/j.ajmo.2025.… EASD 2026 programme easd.org/uploads/EASD2026_Fi… @EASDnews @AmDiabetesAssn @HelmholtzMunich

Sources and notes The main source is Zhao S, Kuo CL, Diniz BS and colleagues, "A Deep-Learning Based Biomarker of Systemic Cellular Senescence Burden to Predict Mortality and Health Outcomes", Aging Cell, online 25 September 2026: doi.org/10.1111/acel.70737. It was a medRxiv preprint first. Funding was NIH grants only, with no conflicts declared. A neural network combines the 38 proteins into one score. It's a proxy for the inflammatory output of senescent cells, not a count of cells, and the authors write that the plasma sources of these proteins are unknown. The 41% is an adjusted hazard ratio of 1.41 per standard deviation (95% CI 1.25 to 1.60), mean follow-up 13.3 years (the tweet's about 13), in the held-out UK Biobank test set (7,639 people, 695 deaths). It compares death rates at any given moment, not lifetime probability. The exercise result is from MEDEX (Mindfulness, Education and Exercise), NCT02665481, a 2x2 factorial trial in older adults with subjective cognitive concerns: clinicaltrials.gov/study/NCT… 585 were randomised and 475 (81.2%) completed 18 months. Its primary cognition result was null (Lenze et al, JAMA 2022): doi.org/10.1001/jama.2022.21… Over 18 months it rose by 0.211 in people not assigned to exercise (that group got mindfulness or health education, not nothing) and by 0.075 in those assigned to exercise (95% CI -0.032 to 0.181, p=0.354). No group or group-by-time effect was detectable, so the gap is too small to rule out chance. The abstract's claim that exercise clearly changed the trajectory isn't supported by its own statistics. A sceptical GP would say this is observational, so the score may partly reflect general inflammation or tissue damage rather than senescent cells. It doesn't show that lowering the score helps. @uconnhealth @washumedicine @ucsandiego @uk_biobank

Sources and Notes The new result. Monte Rosa Therapeutics, GFORCE-1 phase 1, company release 1 Oct 2026 (not peer reviewed): biospace.com/press-releases/… Registry: clinicaltrials.gov/study/NCT… (the registry still lists an estimated 100 for the whole three-part study; the release reports 108 in this part) Design: 108 adults with obesity and raised cardiovascular risk, randomised to MRT-8102 5, 20 or 40 mg once daily or placebo, 4 weeks of dosing and 4 weeks of follow-up. The seven markers in the post are median falls on the drug: hsCRP 85%, IL-6 54%, calprotectin 56%, SAA 51%, S100A12 46% (the three proteins the company links to plaque), fibrinogen 28%, Lp(a) 24%. HMGB1 fell 40% and IL-1beta 37%, but those figures are only for people in the top quarter at baseline, so they're left out. Results were similar at all three doses. Median NEK7 breakdown was about 80 to 90%. The release doesn't say whether the falls are placebo-adjusted and gives no confidence intervals. LDL and triglycerides didn't rise. Safety: no serious adverse events; adverse events in 33% on drug vs 30% on placebo; no sign of more infections over 4 weeks. Next: GFORCE-2, a 6-month phase 2b study of artery imaging and blood markers in stable coronary artery disease, planned for the first half of 2027. No heart-attack outcome trial yet. Why NEK7. The NLRP3 inflammasome needs NEK7 to switch on and release IL-1beta, which canakinumab blocks. IL-1beta drives IL-6, which ziltivekimab blocks, and IL-6 drives CRP; in CANTOS, canakinumab lowered IL-6 as well as hsCRP (Ridker et al., Eur Heart J 2018) doi.org/10.1093/eurheartj/eh… . NEK7 works as a physical part rather than an enzyme, so the protein has to be removed: He et al., Nature 2016 doi.org/10.1038/nature16959 The 2.7 times. Ridker et al., Lancet 2023, 31,245 people on statins in three trials. Top vs bottom quarter of hsCRP: cardiovascular death HR 2.68 (2.22 to 3.23). Top vs bottom quarter of LDL: 1.27. "Die of heart disease" in the post means cardiovascular death, which includes fatal strokes. An association, not proof that lowering hsCRP helps. doi.org/10.1016/S0140-6736(2… The 15%. CANTOS, Ridker et al., NEJM 2017: 10,061 people after a heart attack with hsCRP 2 mg/L or more. Canakinumab 150 mg once a quarter cut heart attack, stroke or cardiovascular death combined, HR 0.85 (0.74 to 0.98); the 50 mg dose didn't reach significance. More fatal infections on the drug. The FDA issued a complete response letter in Oct 2018 and Novartis withdrew the EU filing in Dec 2018: pmlive.com/pharma_news/novar… . doi.org/10.1056/NEJMoa170791… The July miss. ZEUS, Novo Nordisk, 31 Jul 2026: ziltivekimab (IL-6 antibody) in over 6,300 people with artery disease, kidney disease and hsCRP 2 mg/L or more. IL-6 and hsCRP fell, but heart attack, stroke or cardiovascular death didn't: HR 0.99 (0.88 to 1.11). More serious infections on the drug. biospace.com/press-releases/… The mixed record includes a success: low-dose colchicine cut heart events in LoDoCo2 (HR 0.69) doi.org/10.1056/NEJMoa202137… and the FDA approved it (Lodoco) for heart disease on 16 Jun 2023 accessdata.fda.gov/scripts/c… ; it then showed no benefit after a heart attack in CLEAR, HR 0.99 doi.org/10.1056/NEJMoa240592… The same pathway, other pills: Ventyx's NLRP3 inhibitor VTX3232 cut hsCRP 78% vs a 3% rise on placebo over 12 weeks in a 175-person study (64% in the full analysis set), 22 Oct 2025 finviz.com/news/200903/venty… . Lilly agreed to buy Ventyx for about 1.2 billion dollars in Jan 2026 sec.gov/Archives/edgar/data/… . So an 85% hsCRP drop is in line with this drug class; the wider marker panel is what Monte Rosa adds. Lp(a): statins don't lower it and can nudge it up (Tsimikas et al., Eur Heart J 2019) doi.org/10.1093/eurheartj/eh… . The PCSK9 drug evolocumab lowered it 27% (FOURIER) doi.org/10.1161/CIRCULATIONA… . On 4 Sep 2026, Novartis's pelacarsen lowered Lp(a) in 8,323 patients and still missed its outcome trial: novartis.com/news/media-rele… hs-CRP testing: the American College of Cardiology's 2025 Scientific Statement on inflammation (Mensah et al., JACC, 29 Sep 2025) recommends broad hsCRP screening alongside LDL cholesterol, in people with and without heart disease. It's a statement, not yet a guideline. doi.org/10.1016/j.jacc.2025.… The limit: 4 weeks, blood markers only, company data. ZEUS and pelacarsen both showed that moving a marker doesn't prove fewer heart attacks.

Sources and Notes Paper: Aware C ... Lin AL, Journal of Cerebral Blood Flow & Metabolism, published online 10 Sep 2026 (accepted manuscript, not open access). Single-arm pilot, no placebo. doi.org/10.1177/0271678X2614… Mizzou release (Lin's quote on women): showme.missouri.edu/2026/miz… Preprint with the sex split, CC BY 4.0: doi.org/10.21203/rs.3.rs-621… Design: rapamycin (sirolimus) 1 mg a day for 4 weeks, cognitively normal adults aged 45 to 65, brain blood flow by MRI (arterial spin labelling). 23 finished: 9 APOE4 carriers, 14 non-carriers. Carriers rose more than 15% in several regions. Non-carriers showed no significant change. In the preprint, each group was tested before vs after on its own; the published abstract reports no direct carrier vs non-carrier comparison. There was no placebo group. Women (the 35% in the post, described there as an earlier analysis): the published abstract doesn't break results down by sex. Lin's release says women carriers improved most. In the 2025 preprint (an exploratory endpoint, 4 women and 5 men), women carriers rose about 30% in the hippocampus and 35% in the cortex and whole brain (up to about 45% in single regions), and men's change was negligible. Women carriers also started with the lowest flow, though not significantly lower. With no placebo group, some of a rise from a low start can be chance. Blood work in the preprint: carriers' lymphocytes fell from 2.2 to 2.0 (thousand per microlitre), still in the normal range. HbA1c went from 5.4% to 5.5%. IL-6 rose in both groups. APOE4 is the biggest common genetic risk for late-onset Alzheimer's; rare inherited mutations (APP, PSEN1, PSEN2) carry more. 15 to 25% of people carry at least one copy (NIA, nia.nih.gov/health/genetics-… ), hence "up to 1 in 4". In white populations one copy raises Alzheimer's odds about 3-fold (odds ratio 3.2) and two copies about 15-fold, less in Black and Hispanic populations. Farrer et al., JAMA 1997: doi.org/10.1001/jama.1997.03… Leaky vessels in cognitively normal carriers: Montagne et al. (USC), Nature 2020: doi.org/10.1038/s41586-020-2… Rapamycin was first isolated from a soil bacterium collected on Easter Island (Rapa Nui), as an antifungal: Vezina et al., J Antibiot 1975 doi.org/10.7164/antibiotics.… . Rapamycin (sirolimus, Rapamune) was FDA-approved on 15 Sep 1999 to prevent kidney transplant rejection: accessdata.fda.gov/scripts/c… Rapamycin extended mouse lifespan: Harrison et al., Nature 2009: doi.org/10.1038/nature08221 Lin's mouse work: rapamycin restored brain vessel function in Alzheimer's mice (2013) doi.org/10.1038/jcbfm.2013.8… and in APOE4 mice (2016) doi.org/10.1177/0271678X1562… Higher flow in midlife carriers (UK, PREVENT-Dementia, ages 40 to 59): McKiernan et al., JNNP 2020 doi.org/10.1136/jnnp-2020-32… and Dounavi et al., JCBFM 2021 doi.org/10.1177/0271678X2110… Same dose in people already impaired: 10 people aged about 74 with mild cognitive impairment or early dementia took 1 mg a day for 8 weeks (open-label). Rapamycin wasn't detected in spinal fluid, cognition didn't change, systolic blood pressure and HbA1c rose, and spinal-fluid markers of brain injury (p-tau181, GFAP, neurofilament light) rose. Gonzales et al., Communications Medicine 2025 doi.org/10.1038/s43856-025-0… Almost two-thirds of Americans with Alzheimer's are women: Alzheimer's Association 2026 Facts and Figures alz.org/alzheimers-dementia/… Next trial: NCT07728175, University of Missouri, led by Ai-Ling Lin, phase 1, sirolimus vs placebo, 225 women aged 45 to 65 (the registry summary says women who carry APOE4; the numbered entry criteria don't list it), estimated start 1 Dec 2026, primary completion end of 2027, main outcome brain blood flow again. clinicaltrials.gov/study/NCT… Also planned: NIBBLE at Cedars-Sinai, APOE4 carriers aged 45 to 65, fasting-mimicking diet or rapamycin. clinicaltrials.gov/study/NCT…

Sources and notes 1. The video: @WilliamWallace, values read from Figure 2 of Ohayon, Carskadon, Guilleminault and Vitiello, Sleep 2004. 65 studies, 3,577 healthy people aged 5 to 102, recorded by polysomnography or actigraphy. Each age is a different group of people, and the figure is bumpy (55-year-olds sit near 25-year-olds), so "24 fewer minutes" compares the 25 and 65 points. In adults, total sleep fell about 10 minutes and deep sleep about 2 percentage points per decade. Screening out illness and sleep disorders made most age effects larger, not smaller. In studies of people over 60 only, the share of deep sleep didn't change, and only sleep efficiency kept falling. doi.org/10.1093/sleep/27.7.1… 2. Magnesium: Held et al., Pharmacopsychiatry 2002. 12 people aged 60 to 80, two 20-day placebo-controlled crossover periods, sleep EEG. Deep sleep 16.5 against 10.1 minutes, with spreads larger than the gap. The dose built up to 30 mmol a day, about 730 mg of magnesium, twice the US upper limit for supplements. I found no repeat. doi.org/10.1055/s-2002-33195 3. Glycine: Yamadera et al., Sleep and Biological Rhythms 2007. 3 g in volunteers with unsatisfactory sleep; it "shortened PSG latency both to sleep onset and to slow wave sleep without changes in the sleep architecture". doi.org/10.1111/j.1479-8425.… 4. Melatonin: Arbon, Knurowska and Dijk, J Psychopharmacol 2015. 16 healthy people aged 55 to 64, single 2 mg prolonged-release dose; whole-night slow-wave activity unchanged, and lower in the first third of the night. doi.org/10.1177/026988111558… 5. Pink noise: Papalambros et al., Front Hum Neurosci 2017. 13 people aged 60 to 84, one night, lab EEG timing the sound. Slow waves grew while it played; whole-night slow-wave activity didn't change; scored deep sleep was 63.9 minutes on sham and 51.5 with sound (p = 0.06); next-morning word recall improved. Consumer headbands weren't tested. doi.org/10.3389/fnhum.2017.0… 6. Blue-blocking glasses: van der Lely et al., J Adolesc Health 2015, 13 boys aged 15 to 17, polysomnography: "Visually scored sleep stages ... were not modified." doi.org/10.1016/j.jadohealth… The 2023 Cochrane review of 17 trials assessed self-rated sleep only. doi.org/10.1002/14651858.CD0… Red light: I found no trial that scored sleep stages. 7. CBT for insomnia: Sivertsen et al., JAMA 2006. 46 adults with chronic insomnia, mean age 60.8, against zopiclone 7.5 mg or placebo for 6 weeks, sleep recorded at home. "Participants in the CBT group spent much more time in slow-wave sleep (stages 3 and 4) compared with those in other groups." doi.org/10.1001/jama.295.24.… If you take a prescribed sleeping pill, don't stop it suddenly; talk to your prescriber. 8. Age or illness: Vitiello, Moe and Prinz, J Psychosom Res 2002. Of 1,619 older volunteers put through health screening, 51 (3.1%) were found to have a serious sleep complaint or disorder; the authors conclude complaints "cosegregate with illness". doi.org/10.1016/s0022-3999(0… 9. Both directions: Ju et al., Brain 2017, 17 healthy adults aged 35 to 65, "disruption of slow wave activity correlated with an increase in amyloid-beta40" doi.org/10.1093/brain/awx148 Mander et al., Nat Neurosci 2013, frontal grey-matter loss "was associated with reduced NREM SWA in older adults" doi.org/10.1038/nn.3324 Himali et al., JAMA Neurol 2023, Framingham, 346 people over 60 recorded twice: each percentage point of deep sleep lost per year went with 27% higher dementia risk, from 52 cases doi.org/10.1001/jamaneurol.2… 10. The measuring caveat: a slow wave only counts toward deep sleep above 75 microvolts, and older brains make smaller waves (Carrier et al. 2011, doi.org/10.1111/j.1460-9568.…), so part of the late drop is how deep sleep is scored. Video: @WilliamWallace

Sources and notes THE EPISODE Moonshots #297, recorded live 25 Sep 2026, @BenLamm with @PeterDiamandis, @alexwg, @DaveBlundin and @salimismail (@moonshots_pod). Timestamps: edits 20:35, synthesis 21:50, Anthropic 41:00, womb 22:33, p53 9:08, traits 19:43, biobanking 38:13. piped.video/watch?v=VHg3oX1y… 1. EDITS The 300-plus edits at over 90% efficiency, and the 1,000 now in testing, are Lamm's figures. No dataset is public. @colossal announced a cell line carrying over 300 edits in Oct 2024. Its peer-reviewed mouse paper reports up to 11 modifications across 7 genes, in @CellCellPress Cell Reports Methods doi.org/10.1016/j.crmeth.202… The 69-edit pig: Anand et al., @Nature 2023 doi.org/10.1038/s41586-023-0… . Tim Andrews lived 271 days off dialysis on an @eGenesisBio kidney, Jan to Oct 2025. 2. WRITING synXVI, the sixteenth and last rewritten yeast chromosome, Jan 2025 doi.org/10.1038/s41467-024-5… . All 16 have not yet been combined in one living cell. 3. AI @AnthropicAI, 23 Sep 2026: Claude agents found ART, an enzyme system carrying CRISPR-like repeats. Its function is unknown, and Anthropic's own scientists ran the lab work anthropic.com/news/claude-di… AI-written phage genomes that worked: King et al., @ScienceMagazine 2026, from @arcinstitute and @BrianHie doi.org/10.1126/science.aec2… 4. GROWING Eight lambs, supported up to four weeks from about day 105 of a 145-day pregnancy, @ChildrensPhila doi.org/10.1038/ncomms15112 . Mouse embryos grown outside the uterus for 6 of their 20 days, with no births doi.org/10.1038/s41586-021-0… Colossal's own page, posted in May and still live, says it is targeting that birth "within the year" colossal.com/colossal-artifi… Asked about a human baby, Lamm said the technology answer and the ethics answer are very different, and gave no date. 5. OTHER SPECIES African elephants carry at least 20 copies of TP53, humans one. Abegglen et al., JAMA 2015 doi.org/10.1001/jama.2015.13… . Lamm said this work is at one of his companies and small-scale so far. He named p53 and the immortal jellyfish, not elephants. 7. BIOBANKING Colossal and @USFWS signed a memorandum on 25 Jun 2026 covering more than 2,300 threatened and endangered species. No federal funds are committed, and no holdings have been reported. READING, THE FIRST LINE $95,263,072 in Sep 2001 to $525 in May 2022, on @genome_gov's modelled cost-per-genome series, which stops there. Sequencer makers now advertise $80 to $200 on their own arithmetic genome.gov/about-genomics/fa… THE CLOSER @wellcometrust funded the Synthetic Human Genome Project with £10M in Jun 2025. The team aims at one full synthetic human chromosome in 5 to 10 years, and says a whole genome will take decades wellcome.org/news/researcher…

Sources and notes The list: Kapsetaki & Tavernarakis, "How to live for centuries", Aging 2026. 101 species with a maximum lifespan of 250+ years, 11 of them animals (3 sponges, 3 tubeworms, a hydra, a flatworm, a jellyfish, the ocean quahog clam and the Greenland shark) doi.org/10.18632/aging.20641… @AgingJrnl Shark: Nielsen et al., Science 2016. Eye-lens radiocarbon dating of 28 females. Lifespan at least 272 years, maturity at least 156 (give or take 22) doi.org/10.1126/science.aaf1… . Maturity by body size: Nielsen et al., PLoS One 2020 doi.org/10.1371/journal.pone… Hydra: Schaible et al., PNAS 2015. 2,256 hydra from two closely related species, cohorts followed for up to 41 years, death rates flat with age doi.org/10.1073/pnas.1521002… . One other species, Hydra oligactis, does age. PARP: Aoyagi Blue et al., iScience 2021. 23 trees against 36 herbs, 61 plant species in all doi.org/10.1016/j.isci.2021.… Heart: Bergmann et al., Science 2009 doi.org/10.1126/science.1164… Germline: Milholland et al., Nat Commun 2017. Per-division mutation rates: about 81x lower in the human male germline and 68x in mouse, against cultured skin cells from one boy and one mouse pup. Eggs weren't measured doi.org/10.1038/ncomms15183 Kirkwood: Nature 1977 doi.org/10.1038/270301a0 and, naming the disposable soma theory, Kirkwood & Holliday, Proc R Soc B 1979 doi.org/10.1098/rspb.1979.00… Caveats: several lifespans on the list are clonal, estimated or extrapolated. Most comparisons in the review cover two or three species. Disposable soma is a theory with mixed support, and fathers still pass on about two extra new mutations per year of age.

Sources and notes The paper: Zhang, Eckmann, Miao, Mahon and Zou, "The Virtual Biotech: A multi-agent AI framework for therapeutic discovery and development", Science, published 17 Sep 2026. Stanford Department of Biomedical Data Science (@StanfordDBDS) with PHD Biosciences, La Jolla. Lead author Harrison Zhang, a Stanford graduate student. Senior author James Zou (@james_y_zou). doi.org/10.1126/science.aeg6… Free preprint of the same work (bioRxiv, 23 Feb 2026, CC BY-ND 4.0): doi.org/10.64898/2026.02.23.… Stanford Medicine (@StanfordMed) news story, 17 Sep 2026: med.stanford.edu/news/all-ne… How it's built: a chief scientific officer agent takes a question, hands parts to specialist scientist agents (statistical genetics, functional genomics, pathways, chemoinformatics, disease biology, clinical data) and combines their answers. Divisions span target discovery, safety, modality choice and clinical development. Humans set the questions and check the work. The authors call it "human-guided". The trial analysis: more than 37,000 "clinical-trialist" agents curated outcomes from 55,984 trials. Stanford says the agents catalogued them in less than a week. The agents then scored each drug target with single-cell RNA data: how specific it is to one cell type, and how switch-like its activity is (fully on in some cells, off in others). Drugs aimed at cell-type-specific genes were 40% more likely to move from phase 1 to phase 2, 48% more likely to reach market, and had 32% lower adverse event rates. Stanford says the pattern held across cancers and brain, heart, kidney and lung diseases. What a scientist would ask: 1. It's a look back at trials that already ended. It shows which targets did better in the past. No drug has yet been chosen this way and then tested. 2. It's an association. Cell-type-specific targets may differ in other ways too (disease area, drug type, how much genetic evidence backs them). I haven't seen the full paper's adjustments, because it's paywalled. 3. The idea was on the table before. In 2024 a Sanger Institute and Sanofi team (Dann, Teichmann and colleagues) found cell-type specificity predicted progress from phase 1 to phase 2, with a less clear link to approval. doi.org/10.1101/2024.04.04.2… 4. Human genetics is the best known target filter. Drug mechanisms backed by human genetic evidence are about 2.6 times as likely to succeed (Minikel and colleagues, Nature 2024). doi.org/10.1038/s41586-024-0… Why target choice matters: about 90% of drugs that enter human trials fail. Lack of clinical efficacy accounts for 40 to 50% of failures and unmanageable toxicity for about 30% (Sun and colleagues, 2022, analysing 2010 to 2017 trial data). doi.org/10.1016/j.apsb.2022.… AstraZeneca's review of its own 2005 to 2010 pipeline put "the right target" first of the five factors behind success (Cook and colleagues, 2014). doi.org/10.1038/nrd4309 B7-H3: working from data available before January 2025, the agents found B7-H3 high on fibroblasts, the connective-tissue cells around lung tumors, and proposed an antibody-drug conjugate strategy, meaning an antibody that delivers a chemotherapy payload to cells carrying the protein. They proposed a strategy. They didn't design a new molecule. Industry's version: ifinatamab deruxtecan (I-DXd), discovered by Daiichi Sankyo and developed with Merck (MSD outside the US and Canada). Its first human trial began in November 2019: clinicaltrials.gov/study/NCT… It got FDA Breakthrough Therapy designation on 18 Aug 2025 for extensive-stage small cell lung cancer after platinum chemotherapy: merck.com/news/ifinatamab-de… In the phase 2 IDeate-Lung01 trial, 48.2% of 137 patients on the 12 mg/kg dose responded (WCLC 2025): onclive.com/view/ifinatamab-… The FDA accepted the application with priority review, with a decision date of 10 Oct 2026. No B7-H3 drug is yet approved for any cancer: merck.com/news/ifinatamab-de… Stanford calls the match "independent, third-party validation". The drug existed years before the agents' cutoff, and the language models behind the agents may have seen reports of it. So read it as the agents reaching the same answer the field had already reached. I-DXd targets B7-H3 on the tumor cells themselves. Third case in the paper: the agents re-analysed a terminated ulcerative colitis trial of a drug against OSMR beta and suggested it failed partly because the trial didn't select patients by OSMR levels. That's a hypothesis, untested. Funding: Knight-Hennessy Scholarship, NIH (T32-GM145402), NSF, Chan Zuckerberg Biohub, Stanford Biomedical Data Science.

Sources and notes The drawing: Jylhävä, Pedersen and Hägg (@HaggSara, @karolinskainst), "Biological Age Predictors", EBioMedicine 2017, Figure 1. It's a concept drawing, not data. Open access under CC BY-NC-ND 4.0 (creativecommons.org/licenses…). doi.org/10.1016/j.ebiom.2017… Gompertz 1825: doi.org/10.1098/rstl.1825.00… UK death rates by age, ONS National Life Tables (2021 to 2023 and 2022 to 2024): ons.gov.uk/peoplepopulationa… The 2019 analysis (Zhang et al., Genome Medicine) used two cohorts of older Scots and clocks the authors built with growing training sets. The most accurate one had no clear link to death. doi.org/10.1186/s13073-019-0… A 2025 comparison found no link between how well a clock guesses age and how well it predicts death. doi.org/10.1038/s43587-025-0… The Dunedin Pace of Aging, 19 measures at 26, 32, 38 and 45, slowest 0.40 and fastest 2.44 (Elliott 2021): doi.org/10.1038/s43587-021-0… DunedinPACE (Belsky @danbelsky, with Moffitt and Caspi at Duke Psychology and Neuroscience @DukePsychNeuro, @eLife 2022): doi.org/10.7554/eLife.73420. It predicted death in two US cohorts, the Normative Aging Study and Framingham, and the authors report it added prediction beyond GrimAge. Calorie restriction, CALERIE (Waziry 2023, @NatureAging): doi.org/10.1038/s43587-022-0… Where DunedinPACE is weaker than it sounds: 1. It was trained on 817 people from a cohort that is 93% white, and 1.0 is that cohort's average at 45. 2. In 699 Italian adults, the starting reading predicted death, but the change over the years did not on its own. The paper says change added prediction for several clocks once the starting value was counted. doi.org/10.1038/s43587-026-0… 3. It gives the same answer when the same blood is run twice, but in 34 people sampled four times in one day under stress it fell into the "poor" reliability range. doi.org/10.1111/acel.70635 4. CALERIE was analysed after the trial, and people cut calories by about 12%, not the 25% planned. The larger COSMOS trial (958 people, 2 years) found no cocoa effect on any of five clocks, DunedinPACE included. doi.org/10.1038/s41591-026-0… 5. eLife's reviewers asked what it adds over GrimAge, an older clock trained on death. 6. TruDiagnostic holds the commercial licence and sells it. Wearable and app ages: I checked company pages and PubMed. Apple announced Health Age on 9 Sep 2026 without publishing how it's built or tested. Apple's separate pulse-wave age paper (2025) linked its gap to self-reported disease and heart events, with no deaths analysed. WHOOP converts published risk figures into years. Oura checked its number against measured artery stiffness, not outcomes. Life insurers: In July, Munich Re posted a review by Dr Tim Meagher (Journal of Insurance Medicine) concluding that epigenetic clocks add minimal value in underwriting, and that value may improve as multi-omic clocks arrive. munichre.com/us-life/en/insi… A 2024 Journal of Insurance Medicine paper (Mills, Long and Philibert) found that no epigenetic measure has yet been built into underwriting: pubmed.ncbi.nlm.nih.gov/3958… FOXO Technologies set out to price life insurance from a saliva DNA test, sold its insurer in 2023 and says it is no longer pursuing the clock licences (10-K): sec.gov/Archives/edgar/data/… What insurers ask: age, smoking, drinking, height and weight, and personal and family medical history. Larger policies add a paramedical exam with blood pressure and routine blood and urine tests. What comes next: organ ages from blood proteins (Stanford @StanfordMed, Nature 2023): doi.org/10.1038/s41586-023-0… and UK Biobank's protein project, 600,000 samples, full data due by 2027: ukbiobank.ac.uk/news/launch-… No aging biomarker has been approved by US regulators for clinical use (Moqri, Cell 2023): doi.org/10.1016/j.cell.2023.…