A colon cancer called resistant answered, and a scan stood in for a catheter

Most colon cancer belongs to a class that first-generation checkpoint drugs were held not to touch; one month of a two-antibody combination before surgery produced major pathological responses in that class. In coronary disease, a CT scan taken before the procedure guided stenting as well as a catheter-based ultrasound probe threaded into the artery. The first generic of the twice-yearly HIV prevention injection cleared a US regulatory step at a price matching daily tablets, for a list of 120 countries and no others. A general-purpose language model read and rewrote another model's weights directly, with no optimiser and no gradient descent, and the authors are careful about what that did and did not add. Fractional flow reserve, the number that decides whether a narrowed artery is worth a stent, is today's medicine concept.

Connecting the dots

Rules keyed to a world the work has already moved. An essay on law written to expire describes clauses "made flexible, often for set time periods, to allow for real-time adaptation or alteration in response to changing threats, climate or technology", with three requirements: "All experimental clauses must have an expiration date, a clear geographical or sectoral boundary, and rigorous evaluation" (Noema). Set that beside a Kenyan county where malaria control worked: "More than 30 years later, severe malaria is rare to the extent that conducting clinical trials of new therapies and interventions for severe disease is now challenging," and vaccine deployment guidance aimed at moderate-to-high transmission now passes the county by (VaccinesWork). One field is writing expiry into its rules; the other has a rule still pointing at the epidemiology of thirty years ago.

What the catalogue's shape owes to the sheet it was written on. Today's book notes that "the original International Classification of Diseases had some 200 diseases not because of the nature of the human body and its problems but because this was the maximum number that would fit the large census sheets then in use" (Bowker and Star). A new transcriptome-regulation catalogue makes the same point from the other side, by changing the population: "Leveraging elevated autozygosity, we found that causal variants are depleted from runs of homozygosity, consistent with purifying selection" (Nature Genetics). Cousin marriage turns a statistical nuisance into an instrument, and variation that outbred cohorts hide becomes countable.

Where the binding constraint is organisation. England's safety investigator reports that insulin "is a high-risk medication and is among the most common causes of harm from medication errors in the NHS", and that "Most inpatient diabetes care is delivered by non-specialists" (HSSIB). In Kaduna, Nigeria, a pharmacy programme claims that after shelving and labelling, patients "can access the prescribed drugs in less than two minutes" (Nigeria Health Watch). Neither story turns on a device.

Basic biology

A sugar-lipid belongs in the machine that moves folded proteins across a membrane

Twin-arginine translocation carries proteins that are already folded through a membrane, which the folded-protein problem makes awkward. The pathway has been credited to a protein translocon plus the proton gradient. A glycolipid turns out to be required as well, and the system has been rebuilt from purified parts.

"MPIase, a glycolipid that has been known to catalyze membrane protein integration, is essential for the TAT system." — eLife

Why it matters: a protein-only account of a transport machine was incomplete in a way only reconstitution from components could expose, which changes what counts as the machine for anyone modelling or engineering this route.

Atlases and datasets

Transcriptome regulation mapped where reference panels are thin, with inbreeding as the instrument

Qatar Biobank participants, with blood RNA sequencing and whole genomes read together, give a quantitative-trait-locus catalogue for a population that existing panels cover poorly. Quantitative trait loci are the variants that shift how much a gene is expressed or which version is made.

"Fine-mapping resolved 29,953 high-confidence, independent causal loci including 24% previously unreported as well as those with effect alleles showing significant divergence across global populations, while 1,371 were absent from public databases." — Nature Genetics

"These findings establish this QTL catalog as a valuable Middle Eastern resource for genome interpretation and genomic medicine." — same

The authors also run the depletion logic in the other direction, reporting loci with recessive deleterious effects on Mendelian disease genes and homozygous loci driving extreme outlier expression.

Why it matters: clinical interpretation of a regulatory variant depends on having seen it in a comparable population, and a measurable share of what this catalogue resolves is missing from the public databases a clinical laboratory would consult.

Methods

A self-glowing reporter that stops reporting the cell's metabolism

Plants engineered to glow without an added substrate can be watched for days in a whole living organism. The catch is in the signal itself.

"Self-sustained luminescence enables non-invasive monitoring of molecular events at the whole-organism level, but its intensity depends on both gene expression and cellular metabolism, limiting quantitative interpretation." — Nature Plants

The fix is the one fluorescence adopted long ago: a second, constitutively expressed luciferase burning the same substrate as a reference channel, so a change in brightness can be attributed.

Why it matters: ratiometric correction was the missing piece in exactly the class of reporter that works in intact organisms over long periods, rather than in dissociated cells for minutes.

Carrying a hospital-trained model to a handheld probe, and the fix that backfires

A fetal head-circumference model trained on hospital ultrasound carts was taken to a low-cost handheld probe in one obstetric clinic, with each handheld image paired against the same day's hospital measurement.

"Input preprocessing, a choice rarely reported, accounted for about a third of that gap" — arXiv

"fine-tuning on two images significantly worsened accuracy" — same

One device, one clinic, one non-expert annotator, and a small number of analysable patients, as the authors state.

Why it matters: the device class here is the one that shows up wherever a trained nurse is meant to stand in for a specialist, and the result says that a handful of local images is worse than none, while an unreported preprocessing choice carries a third of the measured degradation.

Clinical trials

A mismatch-repair-proficient colon cancer answered a combination before surgery

Mismatch repair proficiency describes most colon cancer, and it has been the marker of futility for this drug class.

"Preoperative immune checkpoint inhibitors (ICIs) are active in patients with deficient mismatch repair (dMMR) colon cancer (CC), whereas proficient mismatch repair (pMMR) CC has been consistently considered resistant to first-generation ICIs." — Nature Medicine

Four non-randomised cohorts, fourteen patients each, one month of treatment before surgery. A CTLA-4 antibody alone did nothing in the proficient tumours; adding a PD-1 antibody produced major pathological responses in that arm, and cleared the mismatch-repair-deficient tumours. The authors state their own limits: no randomisation, fourteen per arm, and an almost entirely white population.

Why it matters: a category of tumour described as resistant produced complete responses inside a month, in a design too small to tell anyone what to do next.

CT before the procedure, instead of a probe inside the artery

Intravascular ultrasound means threading an imaging catheter into the coronary artery to guide a stent. The comparator here is a scan taken beforehand.

"A CCTA-guided PCI strategy was noninferior to IVUS-guided PCI with respect to target-vessel failure at 12 months." — PubMed

"target-vessel failure had occurred in 23 patients (4.1%) in the CCTA-guided PCI group" — same

Why it matters: an imaging step performed inside the body was replaced by one performed before it, with the endpoint holding — in a trial whose discussant notes the patients were not the complex ones, and which was funded by a scanner maker and the company selling the CT analysis.

A pill moved repigmentation in vitiligo, and was still moving it at a year

"STOP-V1: 18.9% versus 6.8%; odds ratio, 3.1; 95% confidence interval 1.6–5.9; P = 0.0003;" — Nature Medicine

"Observed F-VASI75 response rates continued to increase without plateau to week 52 in both studies" — same

The endpoint is three-quarters repigmentation of the face at one year, across two trials. Four skin cancers occurred on the drug and none on placebo.

Why it matters: an oral systemic treatment reached this endpoint in a disease managed with creams and light boxes, and the curve had not flattened when the trials stopped reporting, so the one-year number is not the ceiling.

Hospitals

Insulin: safe at home, supervised by non-specialists in hospital

"Insulin is a high-risk medication and is among the most common causes of harm from medication errors in the NHS." — HSSIB

"Specialist diabetes teams are often under-resourced and unable to provide 7-day coverage" — same

A companion report on the same drug finds the device layer uneven and sometimes unusable: a "postcode lottery" in smart pens, continuous glucose monitors and hybrid closed-loop systems, and pen designs that some patients cannot physically operate (HSSIB).

Why it matters: people who dose themselves correctly at home are often stopped from doing so once admitted, and the expertise to supervise them is not present every day of the week.

Who carries the liability when a clinical model is wrong

The American Medical Association's position puts developers and deploying organisations alongside the clinician, who "cannot be held responsible for AI errors that they cannot control." Physicians, developers and deployers "all bear distinct responsibilities for ensuring the tools do not cause patient harm" (Healthcare Dive). The article cites no statute or decided case.

A US hospital chain's growth arrives uninsured

"Almost half of CHS' 3.6% growth in adjusted admissions this quarter came from uninsured patients" — Healthcare Dive

High-margin elective surgery fell at the same time, which the chief executive attributes to patients deferring care for financial reasons, and the quarter turned to a loss.

Medicine concept of the day

Fractional flow reserve: a narrowing is not yet a reason to stent

The picture of a coronary artery does not say what the narrowing is doing to the blood going through it.

"Coronary angiography shows the presence and degree of stenosis but not the functional significance of the lesion (ie, how much blood flows across the stenosis) or whether a specific lesion is likely to be the cause of symptoms." — Merck Manual Professional

So a pressure wire is passed across the lesion and a ratio is read off it:

"FFR is the ratio of maximal flow through the stenotic area to normal maximal flow obtained during hyperemia (most commonly with adenosine); an FFR of < 0.75 to 0.8 is considered abnormal" — same

Above the threshold, the lesion is left alone: "patients with lesions with FFR > 0.8 or NHPR > 0.89 do not seem to benefit from placement of a stent." Leaving them alone was tested directly, and fifteen years later the untreated group had done better: "The rate of myocardial infarction was significantly lower in the Defer group (2.2%) compared with the Perform group (10.0%)", "without signs of late 'catch-up' phenomenon." (Eur Heart J). Using the number to choose which lesions to treat meant fewer stents and fewer events: "The number of stents used per patient was 2.7+/-1.2 and 1.9+/-1.3, respectively (P<0.001)." and "The 1-year event rate was 18.3% (91 patients) in the angiography group and 13.2% (67 patients) in the FFR group (P=0.02)." (NEJM).

Where it stops. A low reading does not establish that stenting the lesion helps the patient feel better: against a placebo procedure, "PCI did not increase exercise time by more than the effect of a placebo procedure" (Lancet).

The computed versions are where today's trial sits: both of its arms entered on a value derived from the CT scan rather than from a wire, and the commentary on it includes a caution that "not all scanners are the same and not all scans are read in the same way", and that in this trial "the population was not that complex" (TCTMD).

Why it matters: the measured physiological reference standard is being succeeded by estimates computed from images, and the trials keep three questions apart — whether the estimate agrees with the wire, whether decisions made on it turn out well, and whether acting on the number helps at all.

Health systems and policy

A generic of the twice-yearly HIV injection, priced like daily pills, for a fixed list of countries

"limits the sale of the generics to 120 high-incidence, resource-limited countries." — Health Policy Watch

"More than 125,000 people in Africa have already started on lenacapavir" — same

The annual price per person is reported as comparable to daily tablets. The regulatory step is a tentative approval, which lets the product be bought by donor programmes without being marketed in the United States.

Why it matters: the economics of a six-monthly injectable now resemble those of a daily pill inside the licensed territory, and the list of countries is the policy instrument.

The pandemic agreement's last annex opens, with the hard parts open

"Now we must complete the PABS Annex so that the Agreement can enter into force." — WHO

"Of course, there are difficult issues still to resolve." — same

The annex governs access to pathogen samples and the sharing of what is made from them: the term on which laboratories in sampled countries get vaccines and diagnostics derived from their own outbreaks.

A national health data platform counted as live where nobody is using it

"29 NHS Trusts supposedly "realising benefits" from the AI-powered federated data platform (FDP) have either said the system is not live or that they are not using any of its applications." — Novara Media

The replies came through freedom-of-information requests. The operator's own count treats a trust as live when the data pipes exist.

Why it matters: a deployment metric that measures connectivity rather than use will report success for a platform clinicians have not adopted.

Ebola in central Africa

"has killed 5,376 people" — Euronews

"has infected more than 10,000 people, with an over 50% fatality rate" — same

The negotiation above is running during this outbreak.

Big in machine learning

One model rewrote another model's weights, without an optimiser

"Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked." — Takehi et al., MIT CSAIL

"With 50 forward passes, MATH-500 rose from 7.9% to 48.1% and GSM8K from 13.8% to 67.1%." — same

A small fraction of a small model's parameters was changed, and much of the gain came from stopping runaway generation rather than from new competence. One model pair, a research blog rather than a refereed paper, and smaller editing budgets mostly erase the effect.

Why it matters: it makes a network's parameters an artefact another model can read and patch, the way an agent patches code, and the authors draw the line at unblocking something already latent.

Startups and deals

Devices built to be repaired where they are used. A Niger company won an engineering prize for monitors that "can be maintained and repaired in Niger", founded by "the first engineer from Niger to participate in and win the prize" (Disrupt Africa). Repairability is the specification, aimed at imported equipment that dies in place.

People and good examples

Thirty years of ward records in one Kenyan county, and the problem of succeeding

"In 1999, there were about 26 malaria-linked hospital admissions for every 1,000 children in the population. Between 2020 and 2024, this had fallen to less than one admission per 1,000 children each year." — VaccinesWork

The series comes from one county hospital tied to a population under continuous demographic surveillance, which is what makes admissions interpretable as a rate at all. Bed nets and a test-and-treat policy get the credit in the article, and drug and insecticide resistance is named as the live threat.

Why it matters: a decades-long local record can show both that something worked and when a place stops fitting the trials and the eligibility rules written for it.

Reads and ideas

Law with an expiry date, already pointed at clinical software. On experimental clauses: "“Experimental clauses” refer to sections and parts of laws that are made flexible, often for set time periods, to allow for real-time adaptation or alteration in response to changing threats, climate or technology." And the medical instance: "An ongoing Health Sciences Authority experiment allows low- to medium-risk AI medical software to be used in public hospitals before formal registration." The objection is reported too: critics argue broad authorisation "risks turning into backdoor-lobbying channels for Big Tech and powerful corporations seeking exemptions from labor, safety and privacy laws" (Noema). No results from the Singapore experiment are given.

Book of the day

Sorting Things Out: Classification and Its Consequences — Geoffrey C. Bowker and Susan Leigh Star (1999)

"Really good, usable systems disappear, almost by definition. The easier they are to use, the harder they are to see." — excerpts

"Each standard and each category valorizes some point of view and silences another." — same

Their sharpest idea is convergence, where a classification stops merely describing: "This blindness occurs by changing the world such that the system's description of reality becomes true." The worked example is medical, in their telling: classify all disease physiologically, and the observation systems, the records and the available treatments all become physiological, after which the classification looks correct. A reviewer's complaint is worth carrying alongside it — that a book about locality is itself written without much located history (Dynamis).

Why it matters: anyone designing a schema that models and agents will later read, write and train on is choosing where the ambiguity lives, and the authors' practical advice is to place the residual category so that its cost does not fall on one disease or one place.

Corners of the web

A radio receiver was already inside the cheap Wi-Fi board. (translated from German) "Radio researchers from Stuttgart have discovered a hidden debug mode in the ESP32's WLAN radio section that turns the chip into a software radio receiver." — original: "Funkforscher aus Stuttgart haben im WLAN-Funkteil des ESP32 einen versteckten Debug-Modus entdeckt, der den Chip zum Software-Funkempfänger macht." (heise). The firmware "captures raw I/Q samples from the chip's own receiver and sends them to your computer over USB or UART. No extra SDR hardware is needed." (GitHub). Transmit was deliberately left out. It does not replace a cheap dedicated receiver, and the tuning range is narrow — but the capability was sitting in a test path the vendor left behind.