June 2023 — Monthly Briefing¶
1. What was read¶
879 research articles from 33 of the 33 journals on the roster.
A further 243 records (22% of the 1,122 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 |
|---|---|---|---|---|
| 513 | 217 | 77 | 293 | 22 |
Largest contributors: Nature 149, Bioinformatics 121, Science 109, JAMA 93, Nucleic Acids Res 81, Lancet 49.
2. Net change on the thread curve¶
| Thread | This month | Cumulative | Status |
|---|---|---|---|
cell-therapy |
+4 / −2 | +41 / −15 | forming |
ctdna-mrd |
+1 | +14 / −9 | splitting |
foundation-models |
+5 / −2 | +14 / −10 | splitting |
immunotherapy-resistance |
+12 | +66 / −13 | forming |
interception |
+7 / −1 | +40 / −4 | forming |
method-credibility |
+1 / −12 | +1 / −80 | crisis |
neoantigen-vaccines |
+3 | +18 / −6 | forming |
neuro-immune-microbiome-axis |
+9 | +41 / −4 | forming |
next-gen-checkpoints |
— | +11 / −6 | splitting |
single-cell-3d-genome |
+1 | +11 / −5 | forming |
suppressive-population-inversion |
+3 / −1 | +11 / −14 | contested |
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¶
Read first from titles, then re-read from abstracts on 2026-09-04; 45% of this month's cards state something no title carries. The strongest item this month is a refutation and is in section 4(a), where it belongs — and the re-read is where most of this month's refutations came from, since a title that hides its verdict hides a negative one about twice as often as a positive one.
The neuro-immune interface runs in both directions inside a single month, and
that is the argument for calling it a mechanism rather than a crowded topic.
37286594 (Nature): activating an
α2-adrenergic receptor triggers immune rejection of tumours — a neurotransmitter
receptor as an antitumour drug target.
37236193 (Cell): the enteric
nervous system is the relay that converts psychological stress into gut
inflammation. Both point neural → immune. Then
37316669 (Nature) points the other
way: dedicated macrophages organize and maintain the enteric nervous system. A
crowded topic produces many papers pushing the same direction, because that is
what a bandwagon is. A single mechanism network produces papers pushing both
ways across the same interface, because the interface is real and either side can
be the independent variable. The corpus already read the immune → neural
direction in February (36640762,
Cell, immunity to the microbiota promoting sensory neuron regeneration), so June
is a repetition of that direction rather than a one-off — and this time it is
simultaneous with the opposite direction, in the same journal.
37289890 (Science, flagged by the
journal with 37289885) closes the
triangle by adding the microbial arm: a microbiota-modulated checkpoint routing
immunosuppressive intestinal T cells into tumours. Neural, microbial and
suppressive-population accounts described as one trafficking mechanism with a
named control point is precisely the claim neuro-immune-microbiome-axis makes,
and the +9 this month is not a count of on-topic papers.
The axis is dosed in patients, twice. 36863483 (Ann Oncol) gives a defined microbial consortium alongside checkpoint inhibitors in patients with advanced solid tumours — the axis tested as a drug. 36827626 (JCO) reads antibiotic exposure before checkpoint inhibition against overall survival in a population-based cohort of older adults, which is the microbiome effect measured at a scale where mouse-model artefact cannot account for it. The standing objection to this thread has been that dense preclinical publication has no clinical readout behind it; this month it has two, at opposite ends of the evidence spectrum.
Interception gets a randomized readout with ten years behind it. 36917758 (JCO): low-dose tamoxifen versus placebo in breast non-invasive neoplasia, holding at ten-year follow-up. This is drug treatment of a precursor lesion, randomized and placebo-controlled and followed for a decade — the thread's claim tested the way a therapy is tested, rather than inferred from a molecular description of a precursor. 37258665 (Nature) supplies the property the strategy needs to generalize: preneoplasia evolving deterministically under stringent selection, i.e. a predictable trajectory rather than a stochastic one, with 37311413 and 37021392 reading precursor states in myeloma and pancreas the same month.
A personalized RNA neoantigen vaccine induces T cell responses in pancreatic
cancer. 37165196 (Nature).
Pancreatic cancer is the hard case for neoantigen-vaccines: low mutational
burden, and therefore the tumour where "a large enough and predictable enough
target space" is least likely to hold. A title reporting T cell stimulation is
not a title reporting clinical benefit, and this is an immunogenicity endpoint —
but it is the scale claim tested where it is weakest rather than where it is
easiest.
One pretrained model transfers to tasks it was not trained on, and two others
predict experiments they never saw.
37258680 (Nature) states transfer
across network-biology tasks directly rather than assuming it, which is the
foundation-models claim in its own words. In the same month,
37154091 and
37166159 (Mol Syst Biol) predict
cellular responses to unseen perturbations and unseen perturbation combinations.
Three results in four weeks on predict-the-experiment. Read them with 4(a) below,
which is in the same month and does not agree with them.
What is thin, and what is absent. next-gen-checkpoints takes no card at all
this month. The checkpoint content exists — TIGIT in lymphoma, CD24/Siglec-10,
HVEM/BTLA, CD73 with radiotherapy — but all of it is preclinical or a phase 1 of
a target that predates the quarter, so none of it moves a claim about a second
generation arriving. agentic-ai takes no card in any of the three months of
this quarter. That is not for want of material about language models: this month
alone carries 37277596 (Nature, on
coding with ChatGPT) and 37347992
(Science, on AI distorting human beliefs), and March carried
36988602 (NEJM), already carded
neutral. Across 3,021 titles in this quarter there is commentary about language
models and no paper in which one uses tools, plans, or acts. The thread is a claim
about agents becoming infrastructure, and commentary about a chatbot is not
evidence for or against it. Recorded as an absence, with the caveat that an
absence read from titles alone is weaker than an absence read from abstracts —
and that caveat has since been discharged: the quarter's 2,349 research abstracts
were read in full on 2026-09-04 and contain no LLM agent either. The near misses
are transformers used as predictors and one Bayesian optimization loop beating
human process engineers at semiconductor design. This is the only absence in the
corpus that has survived a change of instrument.
On the meeting in this month. Section 6 shows 5,757 ASCO 2023 abstracts,
because the meeting ran 2–6 June. The journal papers above are simultaneous with
it, not downstream of it. A journal title carries no submission date and no
presentation history, so it cannot evidence a podium-to-paper interval in either
direction — it cannot show the meeting leading, and it equally cannot show the
meeting trailing. No clinical-practice card was written for any month in this
quarter for that reason. A reader looking at a month that contains both a major
meeting and a month of oncology literature will otherwise assume the corpus
measured something here. It did not.
4. Refutations¶
(a) Landing on a tracked thread¶
TREM2: one target, three directions, one quarter.
37001042 (Blood) reports that
TREM2 acts as a receptor for IL-34 to suppress acute myeloid leukaemia. Eight
weeks earlier, 37081148 (Nat
Immunol, May) nominated TREM2 macrophages as the driver of NK cell paucity and
dysfunction in lung cancer — a population to remove. In the same month as the
Blood paper, 37601967 (Cell
Genom) reports Trem2+ macrophage recruitment predicting failure of cancer
immunotherapy. Three independent groups, three journals, one molecule, and the
therapeutic direction is not the same in any two of them. suppressive-population-inversion
claims that depleting a population identified as immunosuppressive improves
outcomes; what this shows is that the identification does not transfer across
tumour types, so the depletion logic has no basis for assuming it will. This is
the strongest item in the quarter and it is a refutation, which is the usual shape
of the strongest item on this thread. The supporting side is a genre — nominate a
suppressive population, deplete it, watch the tumour shrink — and it is carded
here, at +3 against this one refutation: intestinal Tregs routed into tumours by a
microbiota-modulated checkpoint, conserved Treg transcriptional connectivity read
as a target set, and stromal and therapy-induced macrophage proliferation in
pancreatic cancer. Each of those names one population in one model. None is an
item of this weight, because the weight here comes from three groups disagreeing
about the same molecule.
Corrected 2026-09-04
The last clause read "the supporting side is a genre of target nominations
rather than a set of citable findings, and cannot be carded against it." The
denominator field in threads.yml it was drawn from has been retracted: the
2026-09-04 abstract re-scan carded 58 of those papers across the corpus,
three of them in this month, and the thread now reads +58 / −112 rather than
the empty supporting side it was given for 42 of its 44 months. The genre is
real; "cannot be carded" was a property of the reading brief, which counted
"population X turns out to be protective" as a finding and "we depleted X and
the tumour shrank" as wallpaper. The TREM2 reading is untouched.
37399355 (JITC): an APRIL-based
CAR shows limited efficacy in patients with multiple myeloma, and the title
attributes the failure to the design choice of using a natural ligand as the
binding domain. A negative clinical result that names a generalizable engineering
constraint is worth more to cell-therapy than a positive one against a single
antigen, because it bounds a class of designs rather than one product. It sits
against 37188782 in the same month,
which extends the platform to CAR-NKT cells in neuroblastoma; the thread is
+4/−2 for the month, and this is one of the two on the other side.
Corrected 2026-09-04
This said "the thread is +1/−1", which set the APRIL CAR failure against a single supporting paper and made the month a tie. It is +4/−2: the abstract re-scan added extragenic safe harbours for T-cell engineering, migration control through ST3GAL1, and a CD22 CAR-T toxicity characterisation, and a second refutation on costimulatory domains directing distinct fates of CAR-T dysfunction. The engineering-constraint argument above is unaffected — it does not depend on the month being balanced.
37248297 (Nat Med): translating AI
from development to deployment in healthcare is written up as the hard part, in
the journal that publishes the development papers. foundation-models claims
transfer into clinical use; this separates benchmark performance from clinical
use rather than treating the two as continuous, and it is one of the two
refutations standing against the month's five supports — the cumulative +14 / −10
is what keeps the thread splitting. Its companion
37326960 (Bioinformatics) asks
whether interpretable deep architectures actually improve drug response
prediction — "myth or reality?" — in the same month as the three
predict-the-experiment results in section 3. Enthusiasm and audit are co-present
here, not sequential.
(b) Method credibility¶
37146621 (Lancet Oncol): a review concluding there is no evidence supporting adjuvant abemaciclib in the indication it is already used in. This is an audit of registration-grade clinical evidence, published in the journal that publishes such trials, and it is not alone this month: 36977286 (JCO) re-evaluates pathologic complete response as a surrogate for event-free and overall survival; and 37172579 (Cancer Cell) argues RECIST v1.1 progression is not the crisp object it is treated as. Three audits of clinical endpoint machinery in one month. The escalation this thread tracks — model, pipeline, instrument, assay — reaches the endpoints that decide what a drug is approved on.
A fourth paper on the same machinery runs the other way and is carded supports:
37338877 (JAMA) applies likelihood
ratios to 169 statistically nonsignificant primary outcomes from 2021 trials in
six leading journals, and finds 91% of them favour the null over the prespecified
effect, 69% by more than tenfold. Nonsignificant results are read here as
evidence of lack of efficacy rather than as an absence of evidence — the
inference the field is repeatedly warned off is, in this sample, mostly sound.
Section 2 shows the consequence: method-credibility reads +1 / −12 for the
month and +1 / −80 cumulative, and that +1 is the only support the thread carries
anywhere in the corpus to this date.
Corrected 2026-09-04
37338877 was listed here as the
fourth of "four audits of clinical endpoint machinery", described as taking
apart the inference of no-effect from nonsignificant trial results. It does
the opposite: it tests that inference and finds it holds in 91% of the
outcomes examined. The title — "Evidence of Lack of Treatment Efficacy
Derived From Statistically Nonsignificant Results" — was read as naming a
problem, which is what every other title on this thread does. The card is
supports, and it is why the cumulative stops being +0.
37198491 (Nature): polygenic score accuracy varies continuously with genetic ancestry. A score reported without the ancestry of the person it is applied to therefore carries an unstated, person-specific error — and read with May's 37388906 on imputation reference panels, the same structural problem appears twice in six weeks at two different layers of the same stack.
Two more that did not take cards but belong to the pattern: 37253541 (Genome Res) finds variants with systematic errors by looking at where genotype discovery approaches disagree, and 37063090 (Mol Syst Biol) is a multi-site assessment of reproducibility in high-content imaging. Both are the genre measuring itself, which is the only form of evidence this thread can accumulate.
5. Open items¶
Nothing carried forward.
6. Meetings¶
5,757 abstracts from 1 meeting.
| Meeting | Abstracts | With full text |
|---|---|---|
| ASCO Annual Meeting 2023 | 5,757 | 5,757 |
7. Regulatory¶
| Event | Count |
|---|---|
| trial-first-posted | 464 (of which 416 genuinely oncology) |
| ema-authorisation | 4 |
| fda-approval | 2 |
2 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:
- 2023-06-15 — FDA grants accelerated approval to glofitamab-gxbm for selected relapsed or refractory large B-cell lymphomas
- 2023-06-20 — FDA approves talazoparib with enzalutamide for HRR gene-mutated metastatic castration-resistant prostate cancer
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.