March 2026 — Monthly Briefing¶

1. What was read¶

736 research articles from 32 of the 33 journals on the roster.

A further 291 records (28% of the 1,027 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
470	197	92	239	29

Largest contributors: JAMA 117, Nature 106, Science 79, Nucleic Acids Res 72, Lancet 46, Nat Biotechnol 44.

2. Net change on the thread curve¶

Thread	This month	Cumulative	Status
`agentic-ai`	+4	+26 / −8	forming
`cell-therapy`	+11 / −3	+219 / −70	forming
`ctdna-mrd`	+2	+79 / −24	forming
`foundation-models`	+5 / −3	+145 / −96	splitting
`immunotherapy-resistance`	+11 / −1	+237 / −89	forming
`interception`	+4	+139 / −24	forming
`method-credibility`	+1 / −18	+22 / −519	crisis
`neoantigen-vaccines`	+2 / −1	+83 / −45	splitting
`neuro-immune-microbiome-axis`	+9 / −1	+208 / −21	forming
`next-gen-checkpoints`	+2 / −3	+74 / −52	splitting
`single-cell-3d-genome`	+3 / −1	+82 / −19	forming
`suppressive-population-inversion`	+2 / −1	+55 / −99	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¶

Three negative checkpoint phase 3s in one month, and only one of them is about a new target. BR.31 finds no disease-free survival benefit for adjuvant durvalumab in resected NSCLC at any PD-L1 level (stratified hazard ratio 0.93, P = .64, at 60 months' median follow-up); KEYVIBE-010 finds no recurrence-free survival benefit from adding anti-TIGIT to pembrolizumab as adjuvant melanoma therapy; and LU005 finds no survival benefit from adding atezolizumab to chemoradiation in limited-stage small-cell lung cancer (median overall survival 36.1 versus 31.1 months, hazard ratio 1.03). All three are the same design move — take an agent that works in advanced disease and push it earlier or add it to a backbone — and all three fail. The adjuvant setting is where this corpus should now expect checkpoint combinations to fail by default.

The distinction that had been collapsed here is worth keeping. Durvalumab and atezolizumab are PD-L1 blockade: BR.31 and LU005 test the existing generation in new settings, and however clean their nulls are they cannot bear in either direction on a claim about targets beyond PD-1 and CTLA-4. Both cards are now neutral on next-gen-checkpoints, and KEYVIBE-010 is the one of the three that actually reads out on a second-generation target.

A target was killed by a reagent artifact. 41871876 (JITC) reports that the high BCMA expression reported in AML is an artifact of nonspecific Fc binding, and that anti-BCMA CAR-T barely lyses AML cells. This is the most economical refutation in the corpus: one antibody-validation control removes an entire proposed indication for a platform that already works elsewhere.

Agentic AI produced three genuine supports in a single month. MOSAIC trains 2,498 specialised chemical experts on millions of reaction protocols and returns executable synthesis protocols at a 71% success rate, with experimental realisation of more than 35 new compounds and reaction methodology that was absent from the experts' training; an agentic system for rare-disease diagnosis carries traceable, auditable reasoning; and LUMI-lab, a foundation-model-driven autonomous platform, discovers ionizable lipid chemistry. Two features are worth separating. MOSAIC's number is wet-lab validated rather than benchmark-scored, which is rare on this thread. And the rare-disease system's selling point is that its reasoning can be checked rather than that it works — a different kind of claim from the rest of the thread, and the one that would survive a sceptical reader.

Vaccines reached unaffected people. 41545594 tests the Nous-209 shared-frameshift-neoantigen vaccine in unaffected Lynch syndrome carriers, and 41818359 (Science) reframes KRAS inhibition as prevention in precursor lesions rather than treatment of advanced disease. Both sit at the join between neoantigen-vaccines and interception, and both move the intervention point earlier than any trial this corpus has previously recorded.

4. Refutations¶

(a) Landing on a tracked thread¶

Corrected 2026-08-27

This subsection carried 41512868 as a refutation. The finding is unchanged and the reading below still holds, but what the paper contradicts is a background assumption about B cells, not a claim this corpus tracks. What it demonstrates — lymphoid-structure integrity governing immunotherapy response, and made druggable — is the neuro-immune-microbiome-axis claim itself, and the card now reads supports. It was recorded against the thread it evidences.

41512868 (Cancer Cell): an IGLL5-positive B cell subset in bladder cancer binds high endothelial venules through IGLL5-LTβR, inhibiting non-canonical NF-κB signalling and disassembling tertiary lymphoid structures, which impairs immunotherapy response in genetically engineered and humanised mice; blocking IGLL5 preserves the structures and improves immunotherapy in patient-derived xenograft and pan-cancer models. TLS have been treated across this corpus as a favourable feature to induce, and the B cell compartment as what builds them. Together with the September 2025 finding that Treg- and MDSC-rich TLS mark fatal cervical cancer, the position now is that TLS presence is not a biomarker and TLS induction is not obviously a goal — what matters is which cells are in them, and here that is a specific subset with a ligand that can be blocked.

(b) Method credibility¶

40790276 (Nat Biomed Eng): latent diffusion models memorize patient imaging data, so synthetic medical images are not privacy-preserving. This is filed under method credibility rather than foundation models on purpose — it is not a capability limit, it is a claim that a stated property of a deployed technique is false.

41772176: histology-based molecular biomarker models learn clinicopathological confounders rather than molecular state. Read against 41372473 from December — foundation models in oncology win benchmarks and miss the clinic — this supplies the mechanism for the miss.

41763219 (Cell Syst): CRISPRa activation levels are confounded by shared cellular resources rather than by guide or target biology. Every CRISPRa screen that reads activation strength as a biological quantity inherits this.

Corrected 2026-08-27

This subsection carried 41653917 as evidence that prior cfDNA end-motif and fragmentomic analyses used partial end representations and so biased published classifiers, and read it as the "second consecutive month with a cfDNA-classifier artifact, after February's demographic-bias card". Neither half survives the abstracts. This paper is a method advance, not an audit: 3' ends were overlooked because standard protocols modify them, and single-stranded "2-end" and "4-end" sequencing recovers the native motifs and detects hepatocellular carcinoma at AUC 0.95, rising to 0.97 with 3' fragmentomic methylation. It makes no claim about published classifiers. February's card was also misread — see that month — so there was no two-month run to name. The card is now neutral, filed on ctdna-mrd as a new modality rather than here.

5. Open items¶

Nothing carried forward.

6. Meetings¶

1,760 abstracts from 1 meeting.

Meeting	Abstracts	With full text
USCAP 115th Annual Meeting	1,760	0

Titles only — these publishers deposit abstract metadata without abstract text, so anything read off this month's meetings is read off titles. The 2024 AACR and ASCO books do carry full text.

7. Regulatory¶

Event	Count
trial-first-posted	728 (of which 650 genuinely oncology)
ema-authorisation	6
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:

2026-03-05 — FDA approves teclistamab in combination with daratumumab hyaluronidase-fihj for relapsed or refractory multiple myeloma
2026-03-20 — FDA approves nivolumab with chemotherapy for previously untreated Hodgkin lymphoma
2026-03-25 — FDA approves relacorilant with nab-paclitaxel for platinum-resistant epithelial ovarian, fallopian tube, or primary peritoneal cancer

8. Trade press¶

25 on-topic of 308 items across 3 feeds (n-stat 298, n-aacr-blog 7, n-acir 3). Headlines are selected by subject keyword, not recency — most of what these feeds carry is health policy and industry finance the corpus does not read.

2026-03-31 — Cancer Today’s Spring 2026 Issue: Using AI Chatbots to Help Inform Care Decisions, When to Monitor a Slow-growing Cancer, and More — n-aacr-blog
2026-03-27 — Editors’ Picks, March 2026: Pediatric Leukemia Development, Low-dose Tamoxifen Use, and More — n-aacr-blog
2026-03-27 — STAT+: Pharmalittle: We’re reading about Novo Nordisk boosting consumer focus, FDA approving a gene therapy, and more — n-stat
2026-03-27 — Decades after Vietnam War, research links Agent Orange exposure to MDS blood cancer — n-stat
2026-03-27 — STAT+: FDA approves Rocket gene therapy for rare immune disorder — n-stat
2026-03-26 — Eli Lilly’s deal man, Allogene’s off-the-shelf CAR-T, and Merck’s Terns buy — n-stat
2026-03-26 — STAT+: Denali Therapeutics, Corcept Therapeutics win FDA approvals — n-stat
2026-03-26 — STAT+: Allogene’s off-the-shelf CAR-T is nearing an early but pivotal study readout — n-stat
2026-03-25 — AACR on Campus Taiwan: Providing a New Perspective on Cancer Research — n-aacr-blog
2026-03-25 — STAT+: FDA approves Denali Therapeutics drug for Hunter syndrome — n-stat
2026-03-20 — STAT+: Novartis picks up experimental breast cancer therapy for $2B — n-stat
2026-03-19 — Bloodlines Series: What Is Smoldering Myeloma? — n-aacr-blog
… and 13 more on-topic items.