March 2023 — Monthly Briefing¶
1. What was read¶
704 research articles from 33 of the 33 journals on the roster.
A further 232 records (25% of the 936 this month's harvest returned) carry no abstract and are not research: news, correspondence, editor highlights and obituaries. PubMed types them Journal Article and nothing else, so the publication-type filter cannot see them — across the whole index they are 22.7% of it and 35% of general, and 3,801 of them are JAMA. Every count below is over the full harvest and is inflated by this; the figure above is not.
| General | Cancer | Immune | Bioinfo | Sysbio |
|---|---|---|---|---|
| 480 | 189 | 79 | 169 | 19 |
Volume swing against the median month: bioinfo -40%
Check the source before reading anything into this. Two benign causes account for most of these. bioinfo every January carries the Nucleic Acids Research database issue and runs high. And several journals simply deposit nothing in some months: across the 44 months read, 4 of 33 journals are absent from at least one, worst being Nat Mach Intell (20 months missing, 81 papers in total) and Bioinformatics (4 months missing, 2,905 papers in total). Every journal on the roster contributed this month. Queried directly, PubMed returns zero for those journal-months too — the holes are in the source, reproduced exactly by re-running the harvest. So a swing here is not evidence that a field went quiet, and is not usually evidence that the pipeline broke either.
Largest contributors: Nature 137, Science 92, JAMA 75, Blood 54, N Engl J Med 51, J Clin Oncol 42.
2. Net change on the thread curve¶
| Thread | This month | Cumulative | Status |
|---|---|---|---|
cell-therapy |
+10 / −3 | +22 / −9 | forming |
ctdna-mrd |
+4 / −3 | +9 / −8 | splitting |
foundation-models |
+4 / −3 | +6 / −7 | contested crisis → contested |
immunotherapy-resistance |
+7 / −5 | +29 / −10 | forming |
interception |
+3 / −1 | +14 / −3 | forming |
method-credibility |
−17 | +0 / −37 | crisis |
neoantigen-vaccines |
+3 / −1 | +10 / −3 | forming |
neuro-immune-microbiome-axis |
+7 | +16 / −3 | forming |
next-gen-checkpoints |
+3 / −1 | +7 / −5 | splitting |
single-cell-3d-genome |
+2 / −2 | +7 / −5 | splitting |
suppressive-population-inversion |
−5 | +0 / −12 | crisis |
Status is derived by rule, not asserted. Fewer than 5 cards → too-early; then refutes ≥ 2×supports → crisis; refutes > supports → contested; supports ≥ 2×refutes → forming; else splitting.
3. The evidence that moved it¶
Written in arrears, and re-read from abstracts
2023 was read after 2024, 2025 and 2026, which still caps what it can
support. It was originally carded from titles and journal names alone; on
2026-09-04 every research abstract of the month was re-scanned with the
instrument 2026 was read with. 41% of this month's cards state something no
title carries. This is also the month agentic-ai opens, and it opens on a
commentary: four readers, three of them working from abstracts, found no
LLM-agent system anywhere in 2023-01..2023-09.
A language model trained on sequence alone predicts atomic-level structure.
36927031 (Science) reports evolutionary-scale structure prediction from
a protein language model, without alignments or templates; 36996195, in the
same journal this month, predicts enzyme function by contrastive learning.
foundation-models claims that scale yields capability that transfers across
tasks, and this is that claim as a result rather than as a resource. It is also
qualified inside the same month, below.
Engineered cell therapy gets a controlled comparison off CD19. 36762851
(NEJM) tests a BCMA CAR against standard regimens in relapsed and refractory
myeloma — randomised rather than single-arm, which no previous card on this
thread has been. 36356599 (Nature) makes the personalised arm
manufacturable by replacing the endogenous receptor non-virally. The thread turns
forming on two different kinds of generalisation landing together: a controlled
readout in a new indication, and a production route for a per-patient product.
The microbiome becomes a dosing variable. 36813961 (Nature): a named bacterial metabolite, 3-IAA, determines whether chemotherapy works in pancreatic cancer. 36867675 (Sci Immunol): checkpoint blockade itself induces gut microbiota translocation that augments antitumour immunity outside the gut. One puts a microbial molecule on the causal path of a cytotoxic regimen; the other makes the microbiome a consequence of immunotherapy as well as an input to it.
Resistance sited outside the tumour, measured in patients. 36931243
(Cell) shows the CD8 response to checkpoint therapy organised in the regional
lymph node and disrupted where that node is metastatic. immunotherapy-resistance
claims resistance is being explained by mechanisms outside the tumour cell; this
locates one in a different organ compartment, in human tissue rather than in mouse
models.
Cell-free DNA carries phenotype, not only presence. 36399432 (Cancer Discov) reads nucleosome positioning in plasma to recover the transcriptional programme of advanced prostate cancer; 36399356, in the same issue, uses fragmentomes to detect liver cancer. Lineage information in plasma is what a treatment choice actually turns on, and it is a different use of the specimen than counting mutant molecules.
What is thin: agents. Across the 2,968 records read this quarter there is no
LLM agent doing work in a tracked journal. What exists is a model under
assessment — 36988602 (NEJM) on the benefits, limits and risks of GPT-4
as a chatbot for medicine, and 36735264 (JAMA) grading a chat model's
cardiovascular prevention advice against guidelines. agentic-ai claims agents
become working research infrastructure rather than demonstrations; a chatbot being
appraised in a clinical journal is the vocabulary arriving, and the single neutral
card is all this quarter will support.
4. Refutations¶
(a) Landing on a tracked thread¶
Neutrophils invert a second time, and now the therapy depends on them. 37001503 (Cell) shows T-cell immunotherapies engaging neutrophils to eliminate the antigen-escape variants T cells cannot see, with 37001504 alongside it. In the same month 36921034 (Sci Transl Med) proposes neutralising a neutrophil product — NET-associated ARG1 — to enhance immunotherapy. Two months running, one cell type carries opposite therapeutic prescriptions in the same literature, after February's pair (36706760, 36800412) said the same thing. The recurrence is the finding: what is unstable is not the target but the premise that a population has one sign.
suppressive-population-inversion closes the quarter at 0/−12, and section 2
calls that crisis. It is not one, and the status rule cannot see why. The rule
fires on the ratio of cards, and through this quarter the thread's supporting side
is uncarded: it is a genre — the templated block that nominates a suppressive
population, deletes it, and reverses immune suppression in mice — running at
roughly a dozen to thirty papers a month across the eight oncology journals, and
no reader of these three months wrote a card for one. Twelve cards means twelve
documented inversions against that stream, not twelve failures out of twelve
attempts. Two of the twelve (CD200 in January, CD47 below) are the sharper case,
where the therapy itself requires the molecule the field proposes to block.
Corrected 2026-09-04
Two things. The count was 0/−6 and is 0/−12: the abstract re-scan doubled the
quarter's inversions, adding CD163-high margin macrophages that resist the
CSF1R depletion aimed at them, immature NK cells that promote triple-negative
breast cancer, an IRF1-dependent neutrophil state immunotherapy requires, and
a PD-1 function inside regulatory T cells. And the reason given for
discounting the ratio was stronger than the evidence for it. This paragraph
said the supporting side "is uncarded by construction" and cited threads.yml
for it; that field has since been retracted, because the same re-scan carded
58 of those depletion papers across the corpus, 49 as mechanism cards. The
genre is real and it was carded as soon as a reader was briefed to see it as
a finding. What holds is the narrower claim now made above: through Q1 2023
the supporting side is uncarded, so the ratio is not a success rate.
36918226 (JITC): CAR T cells require their own CD47 to survive in vivo. Systemic blockade of the canonical don't-eat-me signal would act on the transferred product as well as on the tumour, which is a constraint on combining two of the most-pursued modalities in the field.
36343306 (JCO): the high-risk secondary events seen at myeloma relapse
are present at diagnosis, in subclones too small to detect. This lands on two
threads at once. interception needs the state worth acting on to be measurable,
and here it is explicitly not; ctdna-mrd inherits the same limit, because
residual-disease assays are bounded by the same detection floor. Read with
February's four prognostic MRD reports, the quarter's picture of this thread is a
technology that describes risk well and has not yet been shown to change a
decision.
36894710 (Nat Genet): immune selection determines tumour antigenicity
and influences checkpoint response — the explanation of resistance kept inside the
cancer genome, which is what immunotherapy-resistance claims the field is
leaving.
36521103 (JCO): a survivin-directed peptide vaccine carries the phase
IIa readout in glioblastoma, with the journal's own commentary 36623232
("Antigens Matter") beside it and 36948505 (JITC) reporting a vaccine-based
regimen in prostate cancer. Together with 36070539 in January, every cancer
vaccine that reached a clinical readout this quarter targeted shared antigens.
neoantigen-vaccines claims the private mutation space is large and predictable
enough to support personalised vaccines at scale; whatever the preclinical
antigen-discovery literature says — and 36890230 (Nature) resolves
neoantigen-specific CD8 responses in treated patients this month — the clinical
pipeline in this quarter is not built on it.
36973806 (Genome Biol) qualifies the month's foundation-model result directly: sequence-based expression models capture promoter-proximal determinants and mostly ignore distal enhancers. What reads as learned regulatory grammar is local, and the part of the genome carrying most trait variation is the part the models do not see. 36890378 (Nature) asks the same question of scale in general.
(b) Method credibility¶
Seventeen cards this month, and the one that reaches furthest is 36991470 (Genome Biol): the simulators used to benchmark single-cell methods do not reproduce the data they stand in for. A large share of published method comparisons — the evidence by which tools get chosen — rests on a substrate that was never checked, which makes this an audit of the audit layer rather than of any one tool.
Across the quarter the thread takes 37 cards and every one refutes, which its construction very nearly guarantees. The objects, in order of appearance: a therapeutic editing reagent that does not transfer across people, the readable surface of the literature, the regulatory chain that certifies medical devices, a widely used RNA-targeting effector that kills its host, microbial genotyping that moves with its reference, and the simulators underneath single-cell benchmarking. Most of the other thirty-one are of the same kind — undetected on-target damage after CRISPR editing, off-target conversions from published mitochondrial base editors, an antibody-based DNA-modification map that is confounded by secondary structure, structural-variant callers that disagree in trios, TAD callers whose output depends on the caller rather than the data, a p16 assay that misassigns prognosis. But several of them are models: coevolution models that misrepresent epistasis through undersampling, RNA-binding-residue predictors that fail on the annotation type they were not trained on, machine learning that adds nothing to the pooled cohort equations. This thread is mostly about assays, resources and reagents, and not exclusively so.
Corrected 2026-09-04
This section read "thin this month, and the one card is load-bearing", and put the quarter at six cards with the closing line "only one of those is an algorithm and none is a model, which is worth holding against any reading of this thread as being about AI". The month is seventeen cards and the quarter is 37. Both figures came from a title-only pass, and much of what this thread records — a caller that disagrees with itself, an editor that leaves damage behind, a map that dissolves when its antibody is checked — sits in the abstract and not in the title, so the pass that missed it missed it systematically. The point about the thread's objects survives as a proportion; it does not survive as an absolute.
5. Open items¶
Nothing carried forward.
6. Meetings¶
No abstract book published this month. The abstract layer runs 2023-02 to 2026-08 and is seasonal by nature — meetings are annual, so most months are empty here and that is the expected shape, not a gap in the harvest.
7. Regulatory¶
| Event | Count |
|---|---|
| trial-first-posted | 499 (of which 441 genuinely oncology) |
| ema-authorisation | 5 |
| fda-approval | 3 |
3 FDA actions — dated from each notice's own prose, not the table's posting date, which disagrees on 7 of 190 rows and once by enough to move an approval into the wrong month:
- 2023-03-03 — FDA expands early breast cancer indication for abemaciclib with endocrine therapy
- 2023-03-16 — FDA approves dabrafenib with trametinib for pediatric patients with low-grade glioma with a BRAF V600E mutation
- 2023-03-22 — FDA grants accelerated approval to retifanlimab-dlwr for metastatic or recurrent locally advanced Merkel cell carcinoma
8. Trade press¶
No items. The press archive covers 2024-01 to 2026-08; feeds that cannot be paginated backwards contribute only from the first harvest forward.