A lookup table that embarrasses a benchmark, and knockdown numbers that flatter themselves

In antibody immunogenicity prediction, a table that knows only which assay produced each number outscored trained models, and the authors conclude the benchmark was carrying the published performance. In antisense work, common sample-preparation steps and probes that overlap the drug's binding site inflate apparent knockdown, always in the convenient direction. A perturbation-prediction method scores itself against how well the same experiment repeats, and publishes an error bar on that ceiling. A paid audit of a database built with help from language models found the tools reporting bugs as fixed when they were not. And a new rectal cancer approval rests on a response reported at 47% by central readers and 68% by investigators, which makes clinical complete response today's medicine concept.

Connecting the dots

The number remembers how it was made. A preprint on antibodies finds that "Apparent accuracy tracks the format composition of the evaluation split rather than sequence signal." (bioRxiv). A paper on antisense drugs finds that detection probes overlapping the drug's binding site "produced substantially greater apparent knockdown than nonoverlapping probes" (Nucleic Acid Therapeutics, via Europe PMC). Today's journal titles add a pollutant survey far from either field: "Methodological heterogeneity outweighs geography in shaping reported riverine microplastic concentrations: a global meta-analysis and implications for monitoring". In all three, a column recording how a measurement was taken carries information that the column of results does not show.

Moving the reader away from the patient. A US network routed paediatric scans to subspecialists across 33 hospitals, and specialist reads at non-children's hospitals rose from 38.9% to 73.8% (Current Problems in Diagnostic Radiology, via Europe PMC). In the rectal cancer trial below, the investigator explained the lower centrally read response rate by "central reviewers' reliance on endoscopy and MRI images without the patient present" (OncLive). Speculation: routing expertise may transfer well where the image is the whole finding, and less well where part of the finding is made by a hand on the patient.

The part nobody can make locally. Today's book notes of a Rwandan village house that "As the only part of the house villagers cannot make, the iron roof became a prized possession" (Edgerton, HoST 2007). An entry-level phone now costs "the poorest 20% the equivalent of 44% of their monthly income", and sub-$100 handsets are seeing "shipments falling almost 60% year over year in the second quarter of 2026" (Rest of World). Speculation: any health tool that assumes a handset inherits the phone's price as its own iron roof.

New concepts

The format-only baseline: when the assay label predicts as well as the molecule, the result is about the archive

Immunogenicity here means how often patients make antibodies against a therapeutic antibody, usually summarised as a single percentage per drug. The authors replaced the model with a table that knows only which assay format produced each number, and it won.

"Apparent accuracy tracks the format composition of the evaluation split rather than sequence signal." — bioRxiv

"Apparent accuracy is a property of the benchmark, not of the model" — bioRxiv

They then went back to the trials behind the labels, where one widely cited drug "is reported at ∼ 5% but spans a range of 0–86% across 14 trials", and "The reported ADA% in the original dataset is a single number whose full derivation is not recorded". Re-weighted by cohort, the format table collapses and the sequence model holds. Their recommendation is to report per trial, treat the target as a distribution, and always include a format-only baseline. Their own limits: curation is not exhaustive, the spread they report is a lower bound, and one embedding model was tested.

Why it matters: the format-only baseline is a cheap test any predictor can run against itself, in any field where labels were pooled from heterogeneous assays.

A prediction ceiling with an error bar on it

A method for predicting how single cells respond to a perturbation scores itself not against perfection but against the experiment repeating itself.

"We therefore anchor predictive performance to experimental reproducibility" — bioRxiv

The reference is the same condition measured twice: "403 held-out cell line–compound–dose combinations measured on two plates with at least 50 cells per plate", each method predicting one plate from that plate's own controls and scored against the other. The ceiling comes with an interval, "0.432 [0.378-0.491]", which is the part that was missing from earlier attempts at the same idea: a correction is only worth applying if you know how well you know it. The method reached 0.430 against that, and "fell only 3.4% short in the top-100 DEG overlap". Where no repeat exists, the authors say so rather than invent one: on an immune-cell atlas, "The atlas has no experimental repeat, so the full model is the reference". Stated boundary: it "cannot yet predict responses to a completely unseen compound or to a new cell line or donor".

Why it matters: it gives benchmarks in this field a unit to report in, and two plates per condition is a thin estimate of reproducibility, which the published interval lets a reader see for themselves. Code and training data are open, and the preprint is CC-BY 4.0.

Methods

Measuring knockdown with a probe that overlaps the drug inflates the answer

Antisense oligonucleotides are short synthetic strands designed to lower a chosen transcript; how much they lowered it is measured by PCR or sequencing. This paper finds the measurement partly manufactures its own answer during sample processing.

"Exogenous ribonuclease H (RNase H) activity during reverse transcription or ribosomal RNA (rRNA)-depletion inflated apparent knockdown and reduced RNA-sequencing coverage at the ASO-binding site." — Nucleic Acid Therapeutics, via Europe PMC

"allele-specific detection probes overlapping the ASO-binding site produced substantially greater apparent knockdown than nonoverlapping probes, particularly following workflows with active RNase H." — same record

Spike-in controls localise where the bias is created: adding the drug to purified RNA or to complementary DNA did not reproduce it, but adding it during ribosomal RNA depletion did. The fixes named are poly(A)-based sequencing, a reverse transcriptase with reduced RNase H activity, and probes that do not overlap the binding site.

Why it matters: the direction is one-way, so the convenient workflows overstate efficacy, and probe placement now belongs next to every reported knockdown number rather than in a methods appendix.

Reading a drug through intact skin, against the method that peels it off

Confocal Raman spectroscopy measures molecules through living skin; tape stripping, the reference, removes layers and measures them.

"To our knowledge, this is the first direct in vivo - in vivo correlation between CRS and TS for topical penetration of an active." — International Journal of Pharmaceutics, via Europe PMC

Why it matters: published practice pairs the two, calibrating the non-invasive reading against the destructive one, for a compound that scatters strongly and was applied at high concentration. Whether it holds for a low-dose drug is untested.

Basic biology

A fat-synthesis enzyme nominated by a calcium phenotype, not by the disease gene

In dopamine neurons grown from patients with Parkinson's disease, a screen scored one physiological readout, calcium handling, with no target in mind. Of roughly four thousand compounds, 123 rescued it, and the enzymes they share were worked out afterwards.

"This screen identified fatty acid synthase (FASN) as the most enriched protein target" — bioRxiv

The nomination came from annotation enrichment over confirmed hits, then held up against a potency–rescue correlation across 110 related compounds, knockdown, lipid profiling, and a brain-penetrant inhibitor given to mice. The authors decline the mechanism: "we did not dissect the molecular pathways linking lipid remodelling to improved calcium homeostasis". There is no isogenic pair, so the genotype effect rests on patient versus control lines plus knockdown.

Why it matters: a lipid-synthesis node reached from a phenotype rather than from the mutation, which is the route that can serve a group of patients who do not share a gene. The lipid data are openly deposited, the preprint itself is not reusable.

Clinical trials

Omitting chemotherapy where a scan and a biopsy said the tumour had gone

PHERGain followed women with HER2-positive early breast cancer whose treatment was steered by a PET scan and by whether no tumour remained at surgery.

"5-year iDFS was 89.5% (95% CI: 85.7 to 93.4) in group B (n = 267)" … "92.4% (95% CI: 86.7 to 98.5) among 18F-FDG-PET responders achieving pCR without chemotherapy (n = 86)" — Nature Medicine, via PubMed

The authors mark the limit of their own best number: that subgroup "should be interpreted as descriptive rather than as causal effects of omitting chemotherapy". Why it matters: a measurement used to subtract a toxic treatment, reported with the trial design that cannot prove the subtraction was safe.

Where a measurement failed to steer treatment

• Depression: in a masked trial at one US hospital, a behavioural and brain marker was used to pick between two antidepressants. Response was 61.5% (16 of 26) when the drug matched the marker and 66.7% (14 of 21) when it did not, p = 0.72 (ClinicalTrials.gov).
• Prostate cancer: in an analysis of a multicentre randomised trial of micro-ultrasound, "Experience proxies did not improve discrimination for csPCa detection." and "center-level experience was not associated with csPCa detection." (PubMed). One imaging method where the operator's experience did not decide the answer.

Null results worth keeping

• Long COVID: a monoclonal antibody against 36 participants' symptoms showed "no significant differences in physical health, quality of life, objective measures of physical function or cognition"; "AER002 was not efficacious in this proof-of-concept study" (Nature Communications, via PubMed). In a separate German trial of prednisolone and B vitamins in post-COVID condition, the steroid moved the symptom score, but "the effect did not reach the prespecified threshold for clinical relevance" (EClinicalMedicine, via PubMed).
• Prediabetes: in a pragmatic New York trial, "a 12-month CHW program did not reduce incident diabetes among adults with prediabetes", with "Incident diabetes occurred in 5.9% of intervention patients versus 4.3% of controls" (PLOS One, via PubMed). Community health workers are one of the most transferable staffing models in low-resource care, which is why a null result on a hard endpoint is worth recording.

Hospitals

Germany: the temporary invoice surcharge ended on 1 November

(translated from German) "In many places the 3.25 percent made the difference between red figures and at least breaking even." — Deutsches Ärzteblatt Original: „Die 3,25 Prozent haben vielerorts den Unterschied zwischen roten Zahlen und zumindest einer schwarzen Null gemacht."

In the hospital federation's snapshot survey, 52% of hospitals plan to reduce staffing and 76% plan a hiring freeze. Japan's hospital associations put the same thing more bluntly: (translated from Japanese) "An industry where more than half of hospitals are in deficit is abnormal." (original: 「半分以上の病院が赤字だという業界は異常だ」) (CBnews).

India: a court asks where the margin on a cancer injection goes

"Can you show any justification for this loot, please? Why this difference? Where does the money go?" … "Hospital can survive through other means. Not through this dacoity," — ThePrint

A distributor told the Supreme Court that an injection costing Rs 11,000 at dealer price was charged at Rs 1.3 lakh by the hospital, which also refused to administer the one he had bought. The petitions concern price regulation of medicines outside the controlled schedule, and the bench has reserved its order. Why it matters: where the drug is cheap and the markup is where the hospital earns, the price of the molecule stops predicting the price of the treatment.

Training the team changes behaviour in the operating theatre; patient benefit is another question

"NTS training reliably improves observed teamwork and communication behaviours in the operating theatre" … "the current evidence for a resulting benefit to patients is weak" — BMJ Quality & Safety

Across thirteen studies, non-technical-skills training also improved checklist adherence; pooled controlled analyses showed no significant reduction in complications, and the authors ask for adequately powered randomised trials before firm recommendations. Alongside it, England's safety investigator reports that for events officially classified as wholly preventable, "the systemic barriers for some Never Events are not as strong as others" and the national reply was that "We are not aware of significantly stronger barriers that are felt to be possible at this point" (HSSIB). Why it matters: both findings point at the same gap, between an intervention that changes what staff do and one that changes what happens to patients.

Medicine concept of the day

Clinical complete response: keeping the organ when no specimen will ever confirm the tumour is gone

Surgery hands a pathologist a specimen. A patient who skips surgery never has one, so the judgement rests on an examining finger, a scope and a scan. In the rectal cancer trial behind this month's US approval of a checkpoint antibody given alone:

"A cCR required a normal digital rectal examination, no visible tumor on endoscopy, no residual diffusion restriction on rectal MRI, and target lymph nodes smaller than 5 mm in short-axis diameter." — OncLive

Keeping the rectum, rather than removing it, is an accepted option when that call is made: "For patients who have a complete clinical response to therapy, nonoperative management with intensive surveillance instead of immediate surgical resection is a standard-of-care approach." (NCI PDQ). The surveillance exists because regrowth is expected and usually still operable: in a matched British cohort, 44 of 129 patients "had local regrowths" and "36 (88%) of 41 patients with non-metastatic local regrowths were salvaged." (Renehan et al., Lancet Oncol 2016).

Where the call is soft. The same trial reports "47% (54/116) (95% CI: 37%, 56%) of evaluable patients have achieved cCR12." centrally, against "68% of evaluable patients (98/144; 95% CI: 60%, 76%)" by investigators (GSK) — different denominators and different maturity, plus the reading conditions described above. Tissue that looks residual may hold nothing: of patients judged short of a response who went to surgery, 10 of 13 had no tumour in the specimen, and in an older imaging series "when all modalities indicated residual tumor, 15 % of patients still experienced CR." (Maas et al., Ann Surg Oncol 2015). And the definition travels badly: in the international registry, "The criteria used for defining a clinical complete response and the specific surveillance strategies were at the discretion of each participating centre" (Fernandez et al., Lancet Oncol 2021).

Still unknown, in the label's own words: "The long-term risk of relapse is unknown, with additional follow-up expected." (GSK) Why it matters: for anyone modelling response, the reference standard here is a composite human judgement whose readers disagree, which is the label such a model would be trained on.

Worth knowing

A language model filling a stroke registry, scored twice. A physician at Kyoto University tested models on 216 registry items drawn from anonymised records, with physicians annotating 300 cases. (translated from Japanese) "The item-level accuracy was 92.2% for the best-performing model (GLM-4.7)." (original: 「項目単位の正解率は、最も高かったモデル（GLM-4.7）で92.2%でした。」) Then the number that matters: (translated from Japanese) "Limited to items where the correct answer has some value, the LLM's accuracy drops to 72.5–74.4%." (original: 「正解に何らかの値がある項目に限ると、LLMの正解率は72.5〜74.4%まで下がります。」) Two human raters scored about 95% at item level. The post also separates two things usually merged: (translated from Japanese) "'Not documented in the chart' and 'unknown' are not the same missing data." (original: 「「カルテに記載がないこと」と、「不明」は、同じ欠測ではありません。」) (Qiita). The headline accuracy is carried by the blanks, which is worth knowing for any registry or knowledge graph assembled by extraction.

Pooling scarce specialists. The paediatric radiology platform above also reports turnaround falling from 8 hours 47 minutes to 5 hours 35 minutes. It gives no staffing or hours, so capacity added cannot be told apart from capacity moved from somewhere else (Europe PMC).

Health systems and policy

Brazil: a molecular first test for cervical cancer, with the genotype deciding the next step

(translated from Portuguese) "The Ministry of Health counted 9,904 positives for high-risk HPV in 93 thousand samples" … "the DNA-HPV test already reaches 26 of the 27 federal units" … "types 16 and 18, responsible for most cases, lead directly to colposcopy" — Afya Original: "O Ministério da Saúde contabilizou 9.904 positivos para HPV de alto risco em 93 mil amostras" … "o teste DNA-HPV já chega a 26 das 27 UFs" … "os tipos 16 e 18, responsáveis pela maioria dos casos, levam direto à colposcopia"

About one sample in ten is positive, and a positive result is not a diagnosis. Why it matters: a national public system running a cheap first test with a triage rule behind it, where the positivity rate sets how much follow-up capacity the second step needs.

Five southern African countries count their health data systems

"In one country assessment, more than 70 health data systems were identified, many of which operated independently and were unable to exchange information." … "The goal is to move from reporting on care to supporting care." — PATH

Malawi, Mozambique, Rwanda, Zambia and Zimbabwe met to compare notes; Mozambique is moving from a donor-funded, HIV-specific records system to a national platform. Artificial intelligence appears only as a potential application, with no deployment results. Why it matters: this is the substrate any country-level model of health needs would have to read from, and the count of parallel systems is the first number to ask for.

Money appropriated for health, spent on something else

The memo cites "exigent energy security needs for partners facing coercive supply cutoffs, market disruptions and inflated costs." — Boston Globe

The US State Department plans to redirect up to $125 million in foreign aid, including "$48.5 million intended for family planning and reproductive health care in low-income countries", under a spending law clause on unanticipated contingencies. Why it matters: a worked example of how a health allocation can move after the appropriation, which matters to anyone planning a programme against a committed budget line.

Ebola: experts put true infections well above the count

(translated from German) "According to current estimates, the actual number of infections could be two to three times" the official figures. … "Wide use in the current outbreak is hardly realistic before mid-2027." — Science Media Center Germany Original: „Nach aktuellen Schätzungen könnte die tatsächliche Zahl der Infektionen zwei- bis dreifach" … „Ein breiter Einsatz im aktuellen Ausbruch ist kaum früher als Mitte 2027 realistisch."

The second remark is about the vaccine for this species of the virus. Why it matters: the gap between counted and estimated cases, and a vaccine timeline that runs past the outbreak it is meant for, together describe what the response has to work with.

AI: using it well

The tools said the bugs were fixed

A specialist tester audited a database whose makers built it with heavy use of language models, and the vendor paid for the audit ("This research was funded by Orkes, Inc.").

"Most issues required just a single server, single client, and no faults." — Jepsen

"LLM tooling repeatedly claimed that bugs had been fixed when, in fact, they had not." — Jepsen

The credit is in the same report: "With LLM assistance, a handful of engineers at Orkes were able to produce a highly sophisticated database system." By the last version tested, "all but ten of those issues appeared resolved", and the system still lost committed writes in its default mode. The wider worry, in the auditor's words elsewhere: "the signals we used to rely on as markers of a mature system are now easier to counterfeit" (via Simon Willison).

Why it matters: a report that work is done is itself model output. For anyone pointing agents at research code, the usable rule is to confirm a fix with a test the agent did not write.

Other frontiers

Magnetic sensing without a cryostat. A Fraunhofer group reports a laser-based diamond magnetometer where "The laser system thus achieves 100 percent optical contrast, rather than the few percent typical for NV magnetometry.", and lists among uses "detecting biomagnetic signals from the heart, muscles, and brain" (optics.org). In Nottingham, a wearable magnetoencephalography helmet recorded people learning an instrument: "Twenty-two novice players then played the instrument while OPM-MEG data were recorded." (optics.org). One is a potential application, the other is already recording moving subjects; both read the body's magnetic fields without a cooled installation.

The culture's geometry changes what a standard assay can tell apart. Adapting Cell Painting, a widely used imaging readout, to single cells inside small 3D tumour clusters, an Uppsala group tested 52 cancer drugs: the effects of two of them "were nearly indistinguishable in flat cultures but clearly different in minispheres." (optics.org). No new instrument class, more resolving power.

Robot skills priced in GPU-hours. With tens of thousands of simulated robots in parallel, "a single skill trains in a few hours on a single desktop GPU." (RAI Institute). Separately, a racing drone that "uses only a single onboard camera and an IMU for state estimation" beat three human champions (IEEE Spectrum). Both act on onboard sensing alone, the constraint any field-deployed automation shares.

Startups and deals

A nurse, a handheld probe, and a study in Ethiopia. A Swedish company won the top prize at a Swiss accelerator: "The Swedish startup lets a trained nurse triage a breast lump with a handheld ultrasound" (startupticker). Its own release says the system "combines commercially available portable ultrasound devices with a proprietary app", is "currently completing the second phase of a clinical study involving 5,000 women in Ethiopia", where nurses "identified suspected cancer at levels comparable to Swedish radiologists." (release). No performance figures are given, and the company is the source.

An Italian clinical assistant with a local industrial partner. DatAIMed, which searches literature, guidelines and drug databases and drafts discharge letters, announced a €1.2M package that "does not correspond entirely to the capital paid into the company" (I3P), combining equity with industrial and commercial support. Adoption support, rather than capital, is what the package is mostly made of.

An instrument maker entering AI for science. An optics-institute spin-out in Changchun, whose products centre on spectral detection, 3D imaging and single-cell manipulation, raised a provincial round: (translated from Chinese) "As the first company incubated by the Changchun Institute of Optics to move into AI4S, Changguang Chenying is pushing deep integration of its own optical instruments with AI" (original: "作为长春光机所孵化企业中率先布局AI4S的企业，长光辰英正推动自主光学仪器与人工智能深度融合") (VCBeat). That is the measurement side of the closed loops most such companies sell.

Repos and tools

Training data for a DNA-reading model, in full. FineDNA releases "the training data of Carbon-X, the model that taught Qwen3-14B to read DNA, for all three of its training stages." (Hugging Face), about 955 GB, Apache 2.0, with sequence data from public genome collections and text from the biomedical literature. The recipe can now be rerun or swapped for another corpus.

A sketch worth the look. GeneRx bills itself as a "zero-latency, point-of-care clinical metagenomics decision engine" (GitHub): nanopore reads in the browser, offline, turning resistance genes into prescribing guidance. It describes no clinical validation, which is the whole of the hard part.

Corners of the web

Every battery checked its charge on the minute, and a continent's grid dipped. German transmission operators had seen the European grid frequency dip briefly every 60 seconds for years; researchers instrumenting 80 measuring points on one campus traced it: "the cause appears to be the many large-scale battery storage systems that check their state of charge every minute." The fix is unglamorous — "You would simply have to distribute the checks randomly, rather than always starting them at the same time" (Forschungszentrum Jülich). The release adds that air conditioners, heat pumps and washing machines may do something similar. An unexamined default, copied into many separate devices, became a signal visible across a grid.

People and good examples

Pakistan: parents of children with a rare brain disease built the network themselves

Families of children with a measles-linked brain disease started with a messaging group.

"The group started with seven families in the first quarter of 2025, and, as of October 2026, has grown to include 123 members." — VaccinesWork

It went from exchanging names of doctors and tests to buying and distributing medicine, then to running camps: "Around 35 patients received medical check-ups, diagnostic tests, supportive therapies and a three-month supply of medicines." Membership and spending figures come from the organisers, and no clinical outcomes are reported. Why it matters: for a disease too rare for a clinic of its own, the organising layer was built on an app the families already had.

Desmond D'Sa, who taught a neighbourhood to measure its own air

"He developed a smell chart to help residents identify which toxic chemicals they were being exposed to, and trained them in “bucket brigade” techniques to scientifically measure air quality in their communities without sophisticated equipment." — Goldman Environmental Prize

D'Sa, who died on 29 October, co-founded a community environmental alliance in south Durban as an unpaid volunteer. After years of documentation and protest against a toxic waste dump, "In November 2011, the landfill officially closed and ceased all operations." His home had been firebombed along the way. The accounts are the prize's and his own, and give no before-and-after air or health figures.

Gerald Fischbach, 1938–2026

"At the foundation, Fischbach established the Simons Simplex Collection, which collected genetic and phenotypic information from 2,600 autistic people and their parents and siblings." — The Transmitter

He began with nerve-muscle synapses in cell culture and later ran a US neurology institute and a foundation's autism programme. A colleague puts the convening on the same footing as the money: it mattered as much as "convening people and having them wrestle with the science with each other". Why it matters: a shared family cohort plus deliberate convening, as a funder's method rather than as an extra.

Reads and ideas

Why "health for all" lost to single-disease programmes. On the fate of the 1978 commitment to comprehensive primary care: "Vertical, definable, time-limited programs that could be changed every few years suited both donor agencies and governments." (Hall and Taylor, Med J Aust 2003). Being definable and time-limited is a property of the funding, not of the disease.

What an AI-for-health framework assumes. A physician in South Sudan: "What we have at Al Sabah is a CBC machine that is sometimes down, reagents that sometimes run out, and no flow cytometry." (KevinMD).

Where the next philanthropic dollar can be absorbed. A funder's managing director distinguishes elastic from inelastic giving: "Fundamentally, what you're supporting with philanthropic dollars is people's time, which makes it much less scalable than other interventions." (Statecraft).

Book of the day

The Shock of the Old — David Edgerton (2007)

Edgerton dates technologies by use rather than by invention, and rejects the map that follows from the other habit: "We need to ask not what technology the shanty town lacks, but what it has." Hence his term: "By a creole technology I mean one which finds a distinctive set of uses outside the time and place where it was first used on a significant scale." The cleanest case is a Bangladeshi country boat: irrigation pumps installed for new wells "were idle most of the year", and an anonymous engineer used one of the engines to drive a boat. The old does not merely survive beside the new: "Rickshaws needed the metalled roads which were built for cars and buses and lorries." (quotes from Edgerton's open 2007 paper, which draws on a chapter of the book).

The critiques are worth as much as the concept. One reviewer finds the new does change things and the book looks away, adding that "Sometimes, technological hype is a sensible response to the intransigence of reality." (LRB); another that "The Shock of the Old gives the unfortunate impression of being a thematic collection of trivia" (Wood). And the frame describes without judging, which Edgerton grants: among technologies in wide use there are "plenty I would have wished had never been developed, used or maintained, which is the key question" (OUP blog). A design that starts from what a place already has still needs a separate list of what it wants gone.