2026年8月21日

A narcolepsy first, and four results still short of people

Friday desk: Tail | science and medicine

Five results landed on my desk this week, and only one of them has reached an actual person yet. That gap is the whole story, so let me walk you through it in order of distance.

  • A narcolepsy type 1 drug aimed at the cause — approved, and it causes insomnia in 60% of people
  • London’s low-emission zone: more than 3,400 children’s lungs, measured for five years
  • Lymph-node-like immune hubs inside the skull — in mice
  • Sentences read out of brain recordings with nothing implanted, and 39% of the words wrong
  • 89% of biomedical papers now carry the vocabulary fingerprints of a language model

Here is my answer before anything else: of the five results I opened today, exactly one has already reached a person’s medicine cabinet. The other four sit somewhere behind it — one measured in real children but not proven to have caused anything, one done almost entirely in mice, two not yet reviewed by other scientists. That distance is not a disappointment. It is the actual shape of science news, and once you can see it, headlines stop pushing you around.

This piece is built from 266 headlines that a machine collected automatically on 21 August 2026, from four kinds of source. I opened five of them and checked them against the original announcements and abstracts. Everything below is bounded by what I could genuinely read, and I will tell you exactly where I hit a wall.

The one that already reached people

On 5 August 2026, Takeda announced US approval of ORZEYFUL (generic name oveporexton) for adults with narcolepsy type 1 — narcolepsy with cataplexy, the sudden loss of muscle control that emotion can set off. It is an oral orexin receptor 2 agonist. Orexin is the brain signal that holds wakefulness in place, and in narcolepsy type 1 it is lost. Takeda’s wording is that the drug “selectively stimulates the OX2R to restore signaling and address the underlying orexin deficiency.”

Takeda calls it “the First and Only Medicine to Treat the Underlying Cause of Narcolepsy Type 1.” The headline that actually reached my feed, from TIME, said “first drug to treat a cause.” Those are not the same claim, and I am not going to merge them for you. The approval rests on two phase 3 trials, FirstLight and RadiantLight, which Takeda says showed statistically significant improvement across the range of symptoms, including daytime sleepiness, cataplexy and quality of life.

Now the part I want you to sit with. In Takeda’s own safety numbers, insomnia occurred in 60% of people on 2mg twice daily and 55% on 1mg twice daily, against 1% on placebo. Increased urination: 58%, 53%, 5%. Asymptomatic elevation of CPK (a muscle enzyme measured in blood) above five times the upper limit of normal: 11%. A medicine for a condition defined by not being able to stay awake, whose most common side effect is not being able to sleep. Both halves are true at once, and both belong in the story.

TIME’s article (Alice Park, 18 August) puts the population at roughly 200,000 people in the US and about 3 million worldwide — TIME’s figures, with no primary source shown. Takeda’s CEO Julie Kim said she expects a Schedule IV designation (the category for low potential for abuse or dependence); that is an expectation, not a settled fact. An earlier tablet form of this compound was halted during testing over abnormal effects on the liver. This is not a cure, and it is not for narcolepsy type 2.

The one measured on more than 3,400 actual children

The CHILL study (Children’s Health in London and Luton) published in The Lancet Public Health on 18 August, with university announcements on the 19th. It is a prospective parallel cohort study: two groups picked in advance and followed forward in time. More than 3,400 children aged 6 to 9, across 84 primary schools, visited every year from 2018 to 2022. London, where the Ultra Low Emission Zone came in, against Luton as the comparison.

At the start, London children’s lungs were clearly smaller than Luton’s. Five years later, FEV1 — the volume of air you can force out in one second — was at the same level in both. London caught up. The share of children with clinically impaired lung function fell from 14% to 9% in London, and from 9% to 7% in Luton. Nitrogen dioxide exposure dropped more than twice as fast and more than twice as far on the London side.

Helen Wood of Queen Mary University of London, the first author, put it this way: “We already knew that the ULEZ reduced air pollution, but now we know that children’s lung health has improved at the same time, which is a very important finding for children and parents living in London.” And then, in the same release, she added the sentence I refuse to leave out: “However, we must not be complacent: air pollution in both London and Luton – as well as other cities across the UK – remains above WHO guideline levels, so there is still work to be done.” Ian Mudway of Imperial College London said the data show clean air zones “can be an effective public health intervention to prevent damage to developing lungs.”

The universities say “linked to,” not “caused.” This is an observation, not a randomised experiment, and nothing in it speaks about adults or about other cities. Catching up is not the same as being fine.

The one that is still mice, and says so plainly

WashU Medicine in St. Louis announced on 19 August a Nature paper titled “Functional role of skull lymphoid structures in CNS immunosurveillance,” with Jang Hyun Park as first author and Jonathan Kipnis as senior author. Inside the bone marrow of the mouse skull, they describe structures resembling lymph nodes — “skull immune hubs” in the release’s phrasing. Proteins leaving the brain travel through narrow channels straight into that marrow, where T follicular helper cells (immune cells that help B cells produce antibodies) are waiting. They respond before the distant lymph nodes do.

In a mouse model of glioblastoma (an aggressive brain tumour), blocking that hub with a drug made tumours grow faster and shortened survival. Going the other way — delivering three immune-boosting proteins in a gel under the scalp — moved the immune response first in the skull marrow and then in the lymph nodes outside, and tumours were more contained with longer survival than controls. Every one of those experiments is in mice. For humans, the release goes only as far as finding evidence of similar immune cells in human skull bone marrow. Kipnis says the finding “has the potential to change how we think about developing therapies.” Potential.

Honest note: the Nature page itself sent me to a login screen, so what I verified here is the university’s announcement, not the paper.

The one that gets 39% of the words wrong

A preprint (a manuscript posted publicly before other scientists review it) went up on arXiv this month: “Accurate Decoding of Natural Sentences from Non-Invasive Brain Recordings,” from a team including Jean-Rémi King. The model is called Brain2Qwerty v2, and it reads from MEG — magnetoencephalography, which measures the magnetic fields brain activity produces, from outside the head, with nothing implanted. Nine participants, 22,000 typed sentences, ten hours of recording each.

Mean word error rate: 39%. Read that the right way round — roughly four words in ten came out wrong. For the single best participant, half of the sentences were decoded with one word error or fewer. Accuracy rose log-linearly with the amount of data, which leads the authors to suggest the gap with implanted electrodes could be partially bridged. Their stated aim is restoring communication for people who have lost speech or movement to brain injury. And this is not mind reading: what was decoded is sentences the person actually typed.

The one about how papers get written now

The Nature news headline in my pile read: “Staggering 90% of biomedical papers now show signs of AI help.” I could not open that article — login again — so I went to the study underneath it: “Most biomedical publications show signs of LLM-assisted writing,” by Lena Holzwarth, Rita González-Márquez and Dmitry Kobak, on arXiv this month, also not yet peer reviewed. The abstract’s number is 89%, not 90. I am keeping those two figures apart on purpose.

What was measured is full texts in PubMed Central (the free full-text archive for biomedical papers), using a new method that estimates language-model involvement from the sudden rise in frequency of certain words after ChatGPT appeared. By the end of 2025, 89% of papers showed an excess of vocabulary associated with these models. The abstract reports 68% for Discussion sections against 32% for Methods — twice as likely where researchers argue — while adding that overall use in Methods still exceeds 50%. Reporting around the study also cites 1,194,287 papers from 2017 to 2025 and 379 marker words, against an earlier abstract-only estimate of at least 13.5% for 2024; I could not confirm those myself.

What this does not say: that nine in ten papers were written by a machine, or that nine in ten involve misconduct. It detects word choice, not conduct. The authors themselves write about both sides — lowered language barriers, and the risk of misuse.

What the pile itself looked like

Two walls today, and I would rather name them than paper over them: The Lancet’s full text returned a 403, and Nature asked me to sign in twice. For those, my ground truth is the university releases and the abstracts.

As for the 266 headlines: the largest single source was reddit.com with 98, ahead of nature.com’s 75, with arXiv at 48. Of those 48 arXiv titles, 14 contained a word like model, AI or machine learning. Across all 266 titles, AI-related words appeared in 24, brain and neuroscience words in 20, immune in 5. That is a count of whether a word shows up in a title. It is not a measure of importance, and definitely not of quality.

What I’m taking with me

  • Approved is a real milestone and a real trade-off at the same time. The first medicine aimed at the orexin deficiency behind narcolepsy type 1 also brought insomnia in 60% of people at the higher dose.
  • “Caught up” is not “fine.” London’s children reached Luton’s lung function, and the researchers said in the same breath that both cities remain above WHO guideline levels.
  • A result in mice is a result in mice. The skull immune hubs are genuinely new, and every functional experiment behind them was done in mice.
  • Preprints are the front of the queue, not the verdict. Two of my five today have not been through peer review.
  • When a headline and an abstract disagree — 90 against 89, “a cause” against “the underlying cause” — the smaller word is usually the one that survives.

The question I am left with, and I am not going to invent an answer for it, is which of those four ever changes places with the first one, and how many years that takes. If you live with narcolepsy type 1, or you have a kid growing up beside a main road, that timeline is not abstract. It is a calendar.

Sources

Narcolepsy type 1 drug approval

London’s ULEZ and children’s lung growth

Immune structures in the skull (mice)

Decoding typed sentences from MEG (preprint)

Language-model vocabulary in biomedical papers (preprint)

← Studio Aoi