The graph's universal 'every KMT2A fusion loses the SET domain' is a silent, propagating error class
also called false universals in fusion architecture · the every-fusion error
A statement that is true about the famous half of a chromosome swap has been written down as if it were true about all of them. It has travelled into at least three entries and into this beat's own instructions, and nothing in the checking machinery can catch it, because every citation attached to it is correct.
The same thing, for a scientist
A risk entry about an error class, not about a therapy. The statement *'the SET domain and its WRAD contacts are lost in every KMT2A fusion'* is true of direct 5'-KMT2A products and false of two classes: reciprocal der products, expressed in about 45% of t(4;11) infant B-ALL [E1][E2], and 3'-KMT2A rearrangements such as NUP98::KMT2A [E3]. The error is not a citation error - the supporting citations are real and correct - so `verify-citations.py`, `verify-accessions.py` and `verify-claims.py` are all blind to it by construction. It is a **quantifier error**, and its consequences are directional: it makes a whole architectural class invisible, and that class is the menin-inhibitor-resistant one. The risk kind is a compensation/escape risk: the population the universal erases is the population in which the graph's primary therapeutic argument does not apply.
How far it travelled
ana-kmt2a-interactome-systematic-enumerationstates 'WDR5 / RBBP5 / ASH2L — not in the fusion' with no qualifier.subtraction-kmt2a-win-wrad-interfaceframes the interface as absent from the fusion without qualification.subtraction-kmt2a-set-domainopens 'The enzyme part of KMT2A ... is deleted from every leukemia fusion.'- This beat's own dispatch brief states it as a premise: 'the SET domain and its WRAD contacts are C-terminal to the breakpoint and are LOST in every KMT2A fusion.'
And the graph simultaneously holds both counterexamples, in mechanism-reciprocal-kmt2a-fusion-chromatin-opening (which explicitly records that reciprocal products carry the SET domain) and in fusion-nup98-kmt2a (which lists the retained WIN motif and SET domain in its structured data). The knowledge was present and the universal was written anyway, in a different folder, by a different agent, and neither contradicted the other in any way a validator could see.
Why no existing check catches it
The three verification scripts in this repository test citation identity, accession identity and claim-to-abstract overlap. A false universal has a correct citation (the source really does say the direct fusion loses the SET domain), a correct accession, and high claim-to-abstract overlap (the abstract uses the same words). find-contradictions.py looks for the same PMID cited in opposite directions and for confidence-level disagreement, neither of which occurs here.
The detectable signature, if anyone wanted to build one, is a universal quantifier over a class the graph elsewhere enumerates exceptions to. Words like 'every', 'all' and 'never' applied to fusion architecture are checkable against the graph's own fusion entries, because those entries carry structured domains_retained_3 and domains_lost_3 fields. fusion-nup98-kmt2a lists WIN motif (3762-3767) and SET (3829-3945) under domains_retained_3; any entry asserting those are lost from every KMT2A fusion contradicts a structured field in the same repository.
What it would cost to be wrong
- A resistant class stays invisible. NUP98::KMT2A is already documented as menin-inhibitor resistant with reduced HOXA output 3. An analyst who believes all KMT2A fusions share one architecture has no reason to look for it.
- A stratification variable is never proposed. Reciprocal status is a single RT-PCR with a large prognostic effect on conventional therapy 2 and is recorded nowhere in menin-inhibitor eligibility.
- A whole target class is struck on a premise that covers half the cases. The WRAD subunits were struck from this project's target list partly because their interface is lost; in reciprocal-positive patients it is not.
- The error is self-perpetuating. A universal in a summary line is copied into the next entry's prose, and this beat's own dispatch brief is the proof - the coordinator wrote the false universal as an instruction, and an agent that trusted its brief would have reproduced it a fifth time.
| kind | compensation |
|---|---|
| severity | moderate - conceptually systematic, clinically limited to the reciprocal-positive t(4;11) stratum and to a very rare 3'-KMT2A class |
| mitigation | Restate as 'lost from the direct 5'-KMT2A fusion product'. Add a mechanical check: flag any entry using every/all/never about KMT2A fusion domain content and compare against domains_retained_3 across the graph's fusion entries. |
| detectable by existing checkers | false |
| proposed check | grep entries for universal quantifiers applied to KMT2A fusion architecture; cross-check against data.breakpoints[].domains_retained_3 in every fusion node |
affects
[ "drug-revumenib", "drug-ziftomenib", "drug-bleximenib", "drug-enzomenib" ]
entries carrying the universal
[ "ana-kmt2a-interactome-systematic-enumeration", "subtraction-kmt2a-win-wrad-interface", "subtraction-kmt2a-set-domain" ]
entries holding the counterexamples
[ "mechanism-reciprocal-kmt2a-fusion-chromatin-opening", "fusion-nup98-kmt2a" ]
- Correcting the universal could be over-read as 'WRAD is back on the table'; it is not, and the structural and functional arguments against the direct fusion retaining WRAD are unchanged.
- A mechanical universal-quantifier check would produce a large number of true positives that are merely imprecise rather than wrong, and a checker that cries wolf gets switched off.
- The clinically material half of this correction - reciprocal status as a menin-inhibitor stratification variable - would, if acted on, introduce a new eligibility criterion into a paediatric population whose alternative is chemotherapy with an 11.7% four-year event-free survival.
- How many other universal quantifiers in this graph are false? A single pass for 'every', 'all' and 'never' applied to fusion architecture would find out and has never been run.
- Should structured domain fields be treated as the authoritative record, with prose checked against them rather than the reverse? Three of the four entries carrying this universal have no structured domain payload at all.
- Did the false universal originate in this graph or is it inherited from the review literature? If reviews state it this way, correcting the graph does not correct the reader.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The reciprocal t(4;11) product AF4-MLL comprises AF4 exons 1-3 fused to MLL exons 12-37 and therefore retains the KMT2A C-terminus including the WIN motif and SET domain; the AF4-MLL allele is eliminated in about 40% of patients, which worsens outcome.
The role of reciprocal fusions in MLL-r acute leukemia: studying the chromosomal translocation t(4;11). · Wilhelm A, Marschalek R · Oncogene 2021What this evidence cannot show
Construct description from an engineered model designed to represent the infant-typical breakpoint; patient-level exon composition varies, though not in a way that removes the SET domain.
Evidence that argues against this (1)
- Reciprocal expression is partner- and breakpoint-dependent: about half of KMT2A::MLLT1 cases break upstream of MLLT1 exon 1 and express no reciprocal transcript, so the exception cannot be generalised across partners without checking each one. Meyer C et al. 2023
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The reciprocal AFF1::KMT2A transcript was expressed in 45% of t(4;11) cases among 124 Interfant-treated infants and predicted 4-year event-free survival of 62.4% versus 11.7% with significance retained after risk adjustment.
Unraveling the cellular origin and clinical prognostic markers of infant B-cell acute lymphoblastic leukemia using genome-wide analysis. · Agraz-Doblas A, Bueno C, Bashford-Rogers R, et al. · Haematologica 2019What this evidence cannot show
Retrospective, transcript-level, conventional-therapy era, and the effect direction is the opposite of what a second-driver model predicts.
Evidence that argues against this (1)
- If reciprocal expression is a marker of breakpoint position or of clonal history rather than a biologically active second oncoprotein, the architectural exception is real but therapeutically inert, and correcting the universal changes nothing clinically. Wilhelm A et al. 2021
- E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
NUP98::KMT2A arises from inv(11)(p15q23) and retains almost the entire KMT2A open reading frame including the WIN motif and SET domain while deleting exon 1 and thus the menin-binding region; the resulting mouse leukaemia has reduced HOXA output and is menin-inhibitor resistant.
Transforming activities of the NUP98-KMT2A fusion gene associated with myelodysplasia and acute myeloid leukemia. · Fisher JN, Thanasopoulou A, Juge S, et al. · Haematologica 2020What this evidence cannot show
Two molecularly confirmed human patients and one transgenic mouse with 80-week median latency. The menin-inhibitor resistance and HOXA phenotype are recorded in this graph's fusion-nup98-kmt2a entry from this source; only the abstract was read for the present entry.
Evidence that argues against this (1)
- With two confirmed patients and absence from the 160-patient paediatric NUP98 landscape cohort, this class is vanishingly rare, so the clinical cost of erasing it is small even though the conceptual cost is not. Kaltenbach S et al. 2010
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The statement is a quantifier error rather than a citation error: the underlying sources correctly describe the direct fusion, so citation-identity, accession and claim-to-abstract checking cannot detect it.
Leukemia proto-oncoprotein MLL forms a SET1-like histone methyltransferase complex with menin to regulate Hox gene expression. · Yokoyama A, Wang Z, Wysocka J, et al. · Mol Cell Biol 2004What this evidence cannot show
An assessment of this repository's tooling by this beat, not an external finding. It assumes the three verification scripts behave as their own documentation states; they were not re-read line by line for this entry.
Evidence that argues against this (1)
- A structured check is available and was not built: fusion entries carry domains_retained_3 and domains_lost_3 fields, so a universal about domain loss is mechanically checkable against the graph's own fusion records, which means the blindness is a gap in tooling rather than an inherent limit. Meyer C et al. 2023
The error class is the universal quantifier that the two SET-retaining fusion classes falsify.
Names the sentence: 'the enzyme part of KMT2A is deleted from every leukemia fusion' should read 'from the direct 5'-KMT2A fusion product'.
- refutes →First systematic enumeration of the KMT2A fusion interactome, and the diff against this graph
Names the table row: 'WDR5 / RBBP5 / ASH2L - not in the fusion' needs the same qualification.
- related to →The bite-back of striking a target: a wrong strike costs a cure and generates no data that it was wrong
A companion error class: that entry covers wrong strikes, this one covers false universals, and both are invisible to every checker in the repository because every citation involved is correct.