Auditing the seven strikes before applying them: nine of ten decisions stand, four for different reasons than were given, and one clinical argument fails

This establishes identity only, which is precisely the limit `scripts/verify-citations.py` documents. Twelve of the fourteen were read at abstract level and two (26766589, 38989615) in full text; the support judgements in the table above are therefore abstract-level for twelve papers, and an abstract can omit a limitation that the methods section would reveal. No paper was assessed for methodological quality.

• Fetching a record is not the same as reading the right field of it, and this audit's own cycle supplies the demonstration. The registry entry for NCT03666988 was fetched and read here, and E4 was written saying the record 'does not state the sponsor's reason for termination'. Queried through the ClinicalTrials.gov v2 API this session, the same record's statusModule carries whyStopped: 'Overall benefit-risk profile did not support continuation of the study'. The field was always there; the MCP tool used to read the record does not return it. So an audit that verifies identity - the PMID resolves, the NCT resolves - certifies the weakest property a citation has, and the failure mode that actually bit here was depth of read, not identity. Corrected in E4 this cycle. GlaxoSmithKline (sponsor record) 2022

Abstract read, not full text. 'MLL-rearranged ALL' is not resolved to partner in the abstract, so whether KMT2A::AFF1 specifically was among the models cannot be determined without the methods. The combination result is a patient-derived xenograft experiment, not a trial. The point made here is about what the paper concludes, which the abstract states unambiguously, rather than about how well it supports that conclusion.

• The compound in that paper is not a PRMT1 inhibitor, so neither its killing nor its conclusion can be assigned to PRMT1 specifically. Eram et al., who developed MS023, characterised it as an inhibitor of human TYPE I PRMTs as a class: high potency against PRMT1, PRMT3, PRMT4 (CARM1), PRMT6 and PRMT8, inactive against type II and type III PRMTs, protein lysine methyltransferases and DNA methyltransferases, binding the substrate site as shown by a PRMT6 co-crystal, and reducing global asymmetric dimethylarginine in cells. 'MS023 killing parallels baseline FLT3 R972/973 methylation' is therefore evidence about type I PRMT activity in aggregate; the authors' closing sentence about PRMT1 inhibition is an interpretation their tool cannot separate from CARM1 or PRMT6. Eram MS et al. 2016

Full text read (PMC4712026). The screen is a retroviral transduction/transformation assay in mouse c-kit+ HSPCs, which is a colony-formation readout in a mouse primary-cell model, not a human KMT2A::AFF1 system; and the fusions tested are not individually enumerated in the text read, so 'KMT2A::AFF1 was not among the dependent fusions' rests on the two named dependent fusions being MLL-GAS7 and MOZ-TIF2. Absence of KMT2A::AFF1 from a mouse transformation screen is weaker than a negative result in the human disease, and none exists.

• The genetic screen and the pharmacology are not measuring the same thing, which is a reason they can both be right and a reason neither settles the strike. Eram et al. showed MS023 inhibits PRMT1, PRMT3, PRMT4, PRMT6 and PRMT8 with no activity against type II or III enzymes - so a fusion whose transformation survives Prmt1 knockdown may still be sensitive to blockade of the whole type I family, because the remaining type I enzymes that cover for Prmt1 loss are also inhibited. A shRNA screen that finds 'most MLL fusions unaffected' is evidence against a single-gene PRMT1 dependency and is not evidence against type I PRMT inhibition as a strategy; the strike as written conflates the two. Eram MS et al. 2016

Registry record fetched and read; the whyStopped field was re-read directly from the ClinicalTrials.gov v2 API on 2026-09-19. CORRECTED 2026-09-19: this caveat previously read "the record text read does not state the sponsor's reason for termination, so the widely repeated attribution to 'lack of clinical efficacy and dose-limiting toxicities' is not verified from this source and is not asserted here". That was wrong - the field exists and the MCP registry tool simply does not return it. The sponsor's stated reason is 'Overall benefit-risk profile did not support continuation of the study'. It is a composite judgement and does not itself say whether efficacy, toxicity or both drove the decision, so the commonly repeated attribution remains unverified, but the claim that the registry is silent is withdrawn. The trial had no leukemia cohort and no paediatric cohort, so its relevance to KMT2A::AFF1 infant leukemia is indirect in population, disease and compound.

• CORRECTION (2026-09-19). This item previously stated that the registry record 'does not state the sponsor's reason for termination' and declined to assert one. That is withdrawn: the reason is in the record and was missed because the MCP registry tool does not return the whyStopped field. Queried directly through the ClinicalTrials.gov v2 API this session, NCT03666988 carries whyStopped: 'Overall benefit-risk profile did not support continuation of the study', with statusVerifiedDate 2022-02 and results first posted 2022-05-20. The sponsor's stated reason is a benefit-risk judgement, which is neither the 'lack of efficacy and dose-limiting toxicities' commonly repeated nor the absence of a reason this entry asserted; it is a single composite sentence and it does not say which side of the ratio failed. GlaxoSmithKline (sponsor record) 2022

Abstract read, not full text. Three families is a small numerator and the P value comes from the contrast with 1,000 controls rather than from the family count. The finding concerns lifelong germline heterozygous loss in kidney development and is not a statement about transient pharmacological inhibition in haematopoiesis; it is cited here as a caution, not as the reason for the strike.

• CTR9's germline tumour-suppressor role in the kidney runs opposite to what the PAF1 complex does in this disease, where it is a positive, selectively targetable requirement. Muntean et al. used a conditional allele of the PAFc subunit Cdc73 to show the complex is indiscriminately necessary for AML cell proliferation through epigenetic regulation of proleukemogenic targets including MEIS1 and Bcl2, then engineered a dominant-negative MLL fragment that binds PAFc: disrupting the MLL-PAFc interaction selectively inhibited MLL leukaemic cells while sparing E2A-HLF-transformed cells, and in vivo reconstitution assays showed normal haematopoietic stem cell function was unaffected. A strike on a PAFc subunit justified by a Wilms tumour predisposition is therefore struck against the direction of the only leukaemia-specific separation-of-function experiment in the literature. Muntean AG et al. 2013

Abstract read, not full text. A 2002 positional-cloning study; the tumour-suppressor conclusion is stated by the authors as 'suggested' by the data. Same exposure-mismatch caveat as CTR9: germline heterozygous loss over a lifetime is not partial reversible inhibition over months.

• For CDC73 specifically the counter-evidence is not an analogy but the same gene. Muntean et al. deleted Cdc73 with a conditional allele and found the PAFc is necessary for AML cell proliferation via epigenetic regulation of MEIS1 and Bcl2, and that targeted disruption of the MLL-PAFc interface selectively inhibits MLL leukaemic cells without altering normal haematopoietic stem cell function in reconstitution assays. Parafibromin's tumour-suppressor status in hyperparathyroidism-jaw tumour syndrome describes lifelong germline heterozygous loss in parathyroid tissue; the leukaemia experiment describes acute, complete loss in AML cells and a normal-marrow control. Both can be true, and only the second is about the intervention being struck. Muntean AG et al. 2013

Abstract read. A 1997 review, included here as a map of an uncontroversial clinical genetics fact rather than as primary evidence, per the standing preference for primary sources. It is nearly thirty years old; the MEN1 tumour-suppressor status it describes has not been revised since, but any quantitative claim about penetrance should be taken from a modern source.

• The tumour-suppressor-therefore-do-not-inhibit inference has now been tested in humans against this exact protein, and it did not play out as the inference predicts. In AUGMENT-101, 94 patients aged 1.3 to 75 years received the menin-KMT2A interaction inhibitor revumenib; grade 3 or worse events were febrile neutropenia (37.2%), differentiation syndrome (16.0%) and QTc prolongation (13.8%), with a CR+CRh rate of 22.8% and overall response 63.2% in the efficacy-evaluable KMT2A-rearranged population. Whatever the risks of blocking a tumour suppressor's interface, they did not manifest as prohibitive short-term toxicity, and a registration-enabling cohort now exists. Follow-up is far too short to exclude a neoplasia risk, which is the version of the concern that the MEN1 genetics actually supports. Issa GC et al. 2025

Abstract read, not full text. n=11 sporadic tumours plus one MEN1-patient tumour, in lung carcinoid rather than the parathyroid lesions that matter most for the parallel drawn here. Cited to establish the two-hit tumour-suppressor mechanism for MEN1, which is the point the parallel needs; the parathyroid phenotype comes from E7.

• The two-hit mechanism is about biallelic loss sustained over a lifetime in a tissue; the intervention it is cited against is months of interface blockade in another tissue, and the empirical answer for that intervention is now on record. AUGMENT-101 treated 94 patients with revumenib, including children from 1.3 years old, and reported a predictable safety profile dominated by febrile neutropenia, differentiation syndrome and QTc prolongation rather than by anything resembling a loss-of-tumour-suppressor phenotype. The honest residual concern from lung carcinoid and parathyroid genetics is long-latency second neoplasia in survivors, which no trial of this length can address and which nobody is currently measuring in this population. Issa GC et al. 2025

Abstract read, not full text. The arms are small and unequal (35, 36 and 19 patients) and a Simon two-stage design per arm is not a powered head-to-head comparison, so neither the response advantage nor the death imbalance should be read as an estimated treatment effect. The age imbalance is stated in the abstract but no adjusted analysis is reported there, so the confounding is visible rather than quantified. Alvocidib is a multi-CDK inhibitor administered with cytarabine and mitoxantrone, so no observed effect can be attributed to CDK9 or P-TEFb specifically.

• A second randomised trial of the same drug reports the opposite toxicity finding, which is why neither should be used as a general verdict on P-TEFb inhibition. In 165 newly diagnosed intermediate/adverse-risk AML patients, FLAM gave a significantly higher complete remission rate than 7+3 (70% vs 46%, p=0.003) WITHOUT increased toxicity, and with no significant difference in overall or event-free survival. The real objection to this class is that extra remissions did not become extra survivors, which is a very different problem from treatment-related mortality. Zeidner JF et al. 2015

Abstract read, not full text. Mouse, and a CD19-cre B-cell-restricted deletion, so it models complete loss in one lineage rather than systemic partial pharmacological inhibition. It says nothing about leukemia and nothing about KMT2A fusions; it is cited solely to show that the word 'dispensable' is the wrong description of ELL2 and that the strike needs a two-part reason.

• A lineage-restricted conditional-null phenotype in mice is a weak predictor of what a partial inhibitor does in patients, and this disease supplies the worked example. Menin is a bona fide tumour suppressor whose germline inactivation causes a multi-organ neoplasia syndrome, and blocking its KMT2A interface in 94 patients from 1.3 to 75 years old produced febrile neutropenia, differentiation syndrome and QTc prolongation - an ordinary oncology safety profile - alongside a 22.8% CR+CRh rate. The ELL2 objection is right that 'dispensable' is the wrong word for the published mouse phenotype; it does not follow that the phenotype forecasts the clinical liability, in either direction, and the register for both should be 'unmeasured'. Issa GC et al. 2025

Abstract read, not full text. Mouse, MLL-AF9 not KMT2A::AFF1, and concerns wild-type MLL1's SET domain. Critically for the two strikes it supports, the study does not test RBBP5 or ASH2L at all: the strike is an inference that a subunit whose function is to enable a dispensable catalysis is itself dispensable, which is reasonable but is not a result. A demonstrated non-catalytic requirement for either subunit would reverse it.

• This item is the single mouse result on which two of the strikes rest, and a later paper in this same graph reports the opposite about the wild-type allele. Chen et al., using independent conditional animal models, found endogenous MLL1 DISPENSABLE for MLL-rearranged leukaemia; deleting the closest paralogue Mll2 alone significantly impaired survival of MLL-AF9-transformed cells, additional Mll1 loss further reduced viability and proliferation, and the two collaborate by regulating distinct pathways rather than redundantly, leading the authors to nominate MLL2 as the drug target in MLL-rearranged leukaemia. Patricia Ernst is senior author on this and on much of the preceding wild-type-MLL1 literature, so this is the field correcting itself. If the H3K4 methyltransferase that matters in this disease is MLL2 rather than MLL1, an argument that strikes RBBP5 and ASH2L because MLL1's catalysis is dispensable is reasoning about the wrong enzyme. Chen Y et al. 2017

Abstract read, not full text. Entirely in vitro with reconstituted complexes, and a separate entry in this graph notes that reconstituted MLL1 core-complex assembly is strongly temperature- and concentration-dependent with the disassembled state favoured at physiological temperature. In vitro subunit-requirement conclusions may not transfer to the cell.

• Taken together with the dependency data, this in-vitro result inverts the strike it supports. The finding is that Win-motif/WDR5 disruption severely disrupts MLL1 and SETD1A complexes but only MODESTLY disrupts MLL2/4 and SETD1B without significantly altering their activity. Chen et al. then showed that in MLL-rearranged leukaemia the paralogue that matters is MLL2: endogenous MLL1 was dispensable in independent conditional models, Mll2 deletion alone significantly impaired survival of MLL-AF9-transformed cells, and the two regulate distinct pathways rather than substituting for one another. So the complex least disturbed by blocking the WDR5 interface is the complex the disease depends on - which is an argument about which interface to drug, not evidence that the WRAD subunits do not matter here. Chen Y et al. 2017

Abstract read, not full text. The PubMed abstract is garbled where the fusion name should appear ('resistant to leukemogenesis induced by'), so the identity of the driver in the knock-in experiment is taken from the coordinator's pre-verified reading as MLL-AF9 and is not independently confirmed from the abstract text. AML, not infant KMT2A::AFF1 B-ALL. MRT199665 is a tool compound with no clinical development.

• The human half of this evidence is a biomarker-stratified cell-line panel, and cell-line panels are less stable instruments than they appear. Ben-David et al. genomically characterised 27 strains of a single widely used line plus multiple strains of 13 further lines, found extensive clonal diversification driven by positive selection under culture conditions, and - testing the 27 strains against 321 anti-cancer compounds - found that at least 75% of compounds that strongly inhibited some strains were completely inactive in others. A result of the form 'MRT199665 kills MEF2C-phosphorylation-positive lines and spares negative ones' therefore needs its strain provenance stated before the stratification is read as biology, and the redefinition of MEF2C rests on it alongside the knock-in mouse. Ben-David U et al. 2018

Abstract read, not full text. Mouse; complete conditional deletion, not partial inhibition. It establishes that the normal adult stem cell needs Meis1, which is the therapeutic-window objection, but it does not measure whether the leukemia needs more of it than the stem cell does — and that quantitative comparison is the one that would settle the demotion.

• That normal adult stem cells need Meis1 does not by itself close the window, because the window can be opened upstream, and there is a published demonstration of exactly that against a MEIS1-dependent axis. Muntean et al. showed the PAF complex is required for AML proliferation through epigenetic regulation of MEIS1 and Bcl2, and that a dominant-negative MLL fragment disrupting the MLL-PAFc interaction selectively inhibits MLL leukaemic cells, spares E2A-HLF-transformed cells, and leaves normal haematopoietic stem cell function intact in in vivo reconstitution assays. Selectivity came from the interaction, not from the target gene. A MEIS1 demotion argued purely from the normal-HSC requirement is therefore an argument against direct MEIS1 inhibition, not against attacking the MEIS1 axis. Muntean AG et al. 2013

Abstract read, not full text. CML, driven by BCR::ABL1 and not by a KMT2A fusion; K562 is the main model. Relevant to this graph as a demonstration that KDM6A's direction of effect is context-dependent, not as evidence about KMT2A::AFF1.

• KDM6A's direction of effect is context-dependent, and the context closest to this project's disease runs the other way. Van der Meulen et al. found somatic loss-of-function UTX/KDM6A mutations in human T-ALL, exclusively in male patients, showed UTX escapes X-inactivation in female T-ALL lymphoblasts and normal T cells, demonstrated in vitro and in vivo that UTX is a bona fide tumour suppressor in T-ALL, and reported that UTX-inactivated T-ALL shows collateral sensitivity to pharmacological H3K27me3 inhibition. A reframing built on KDM6A promoting imatinib resistance in a myeloid CML line therefore cannot be carried into lymphoid leukaemia without inverting: there, losing KDM6A is the oncogenic event. Van der Meulen J et al. 2015

Abstract read, not full text; no PMC full text is listed. AML broadly, not KMT2A-rearranged AML specifically, and the correlation between KDM6A expression and venetoclax tolerance is an association across samples rather than a demonstrated causal determinant of clinical response.

• The therapeutic logic here and in lymphoid leukaemia point in opposite directions, which is the strongest argument against treating KDM6A as one target with one sign. This item reads KDM6A loss as a sensitiser to venetoclax and PARP inhibition in AML; Van der Meulen et al. found UTX loss-of-function mutations to be recurrent, male-restricted drivers in T-ALL, established UTX as a tumour suppressor there in vitro and in vivo, and reported that the resulting tumours are collaterally sensitive to inhibiting H3K27 METHYLATION - the opposing enzyme. Same gene, same tissue lineage tree, opposite pharmacology, and an infant B-ALL programme sits between the two with no data of its own. Van der Meulen J et al. 2015

Abstract read, not full text. Solid tumours, not leukemia, and the dependency is conditional on SMARCA4 inactivation which is not a common lesion in KMT2A-rearranged infant leukemia. Cited only to support the general point that a near-zero mean DepMap gene effect across unselected lines does not exclude a genotype-restricted dependency.

• The genotype-restricted-dependency argument is symmetric, and its other instance is a warning rather than a licence. This item uses SMARCA4-deficient tumours to argue that a near-zero mean DepMap gene effect does not exclude a restricted KDM6 dependency, which is correct. But the best-characterised genotype restriction for this gene in a leukaemia is UTX INACTIVATION itself: Van der Meulen et al. found recurrent somatic UTX loss-of-function in male T-ALL, demonstrated tumour-suppressor function in vitro and in vivo, and showed those tumours are sensitive to H3K27me3 inhibition. So the same reasoning that says 'a restricted KDM6 dependency may exist' also says a restricted KDM6-loss-driven malignancy exists, and no one has determined which side of that line KMT2A-rearranged infant leukaemia falls on. Van der Meulen J et al. 2015