Oncofusion

start / graph / hyp-partner-outcome-separation-versus-menin-class-assumption

Hypothesisspeculative confidence 1 of 5: Our own inference or analogy. Not shown to be true.Q19-partner-outcome

The clinical outcome data do not support a single KMT2A class, and they do not refute the menin interface either

also called partner outcome versus menin class assumption · Q19 central hypothesis

Patients with different partner genes really do have different outcomes, but only in some age groups and some disease types, and the reasons may have nothing to do with menin. So the outcome data cannot confirm or kill the idea that one drug works for all of them.

The same thing, for a scientist

Statement: partner-stratified outcome separates strongly in children (childhood AML partner risk groups differ by 24 EFS points, n=1,130; non-infant childhood ALL spans 41.7% to 91.2% across partner-by-lineage cells, n=629) and does not separate in infants (two cohorts, n=115 and n=175, both null, the latter at 36% versus 37%, P=0.9) or in the two largest adult AML series (n=172 and n=205, both null). The separation that exists is not uniform in direction: KMT2A::MLLT1 is high-risk in childhood AML and the best-outcome cell in childhood T-ALL. Therefore the class abstraction is neither confirmed nor refuted by outcome data - it is confirmed in the population where menin inhibitors are most likely to be used first in infants, and refuted in children over one year. Crucially, none of the separation observed is attributable to the menin interface, because every fusion in every arm retains the same menin-binding N-terminus. Prior probability that a partner-by-lineage interaction is real and not fully explained by age, protocol era and comutation: 0.45.

Go deeper
What the assembled numbers do say
  1. Partner separates outcome in children. This is robust: a pre-assigned partner risk group independently predicts EFS, CIR and OS in 1,130 children with AML after adjustment for MRD 1, and it has been adopted into COG AAML1831. It is not a statistical artefact.
  2. Partner does not separate outcome in infants. Two cohorts tested it and reported nulls 3, and the protocol that achieved the best infant KMT2A-r EFS did so with age, CNS status and MRD and never used the partner.
  3. Partner does not separate outcome in adult AML, in the two largest series 4.
  4. Where it separates, the direction is lineage-dependent 2.
What they do not say

They say nothing about menin. Every arm in every comparison above carries a KMT2A N-terminus with intact MBM1/MBM2 motifs; the menin interface is held constant across the entire experiment. A partner-outcome difference is therefore evidence that something other than the menin interface differs between fusions - which is compatible with menin inhibition working equally well in all of them, and equally compatible with it not. The audit that dispatched this beat asked whether outcome separation would show the class abstraction to be wrong. It does not, because outcome is not a menin readout.

The reframing this beat proposes

Partner-stratified prognosis was the wrong proxy. The right proxy is partner-stratified response to menin inhibition, and that does not exist in the primary literature - zero reports across five counted queries 5. The cheapest available test of the class assumption is therefore not a new analysis of old trials but the retrieval and publication of per-partner response from the menin trials already run. AUGMENT-101's efficacy population contains 2 KMT2A::AFF1 patients and 13 with no partner identified at all 5, so even that retrieval will yield arms too small to be conclusive, and pooling across sponsors will be required.

What would kill this hypothesis

See data.falsifiers. The sharpest is the first: if the pooled partner-by-lineage interaction term is null with adequate power, then the childhood partner effect is a main effect confounded by age or comutation, the sign flip is noise, and the class abstraction stands on the outcome data.

Facts and figures
statementKMT2A 3-prime partner separates clinical outcome in children over 1 year but not in infants or in adult AML, and where it separates the direction depends on lineage; none of this separation bears on the menin interface, which is held constant across every comparison, so partner-stratified prognosis cannot test the menin class assumption and partner-stratified response to menin inhibition - which does not exist in the primary literature - is the test that would.
origindata-pattern
prior probability0.45
predictions
[
  "A partner-by-lineage interaction term in pooled individual-patient data will be significant.",
  "Adjusting a paediatric partner comparison for RAS-pathway mutation status will shrink but not abolish the partner effect.",
  "AUGMENT-101 Data Supplement Table S3 will show response rates that vary across partners but with arms of 2-13 patients and overlapping intervals, so the variation will be uninterpretable.",
  "Interfant-99 and Interfant-06 per-partner EFS, if retrieved, will be flat, consistent with the two published infant nulls.",
  "Response to menin inhibition will correlate better with lineage and developmental state than with 3-prime partner identity."
]
experiments
[
  "exp-partner-by-lineage-interaction-pooled-ipd"
]
analyses
[
  "ana-partner-outcome-enumeration"
]
falsifiers
[
  "A partner-by-lineage interaction term fitted in pooled individual-patient data with a published power calculation returns null, and the partner main effect disappears after adjustment for age, era and comutation. That would make the childhood partner effect a confound and leave the class abstraction intact on outcome data.",
  "AUGMENT-101 Data Supplement Table S3, or a pooled sponsor dataset, shows CR+CRh rates that differ by partner beyond overlapping confidence intervals. That would refute the claim that outcome separation is uninformative about menin, by showing separation in a menin readout.",
  "A single cohort reports partner-specific EFS for the same partner in both lineages and finds the same direction in each, removing the sign flip that motivates the hypothesis.",
  "A menin-inhibitor trial reports response stratified by partner and finds it flat across arms of adequate size, which would confirm the class assumption directly and make the whole partner-prognosis question irrelevant to the therapeutic bet.",
  "Interfant-99 or Interfant-06 per-partner EFS, once retrieved, shows a large partner effect in infants, contradicting both published infant nulls and the claim that infancy masks partner."
]
How this could bite us back
  • If partner-stratified response is eventually reported and differs, the class label will already have been granted and clinical practice set; the correction would then have to be made against an approved indication.
  • A negative interaction test would be used to justify class-wide menin inhibitor use in populations where the drug has never been evaluated, including the 91.2%-EFS MLLT1 T-ALL group that should probably be excluded.
  • The proposed reframing - measure response, not prognosis - requires sponsor cooperation on subgroup reporting, and the arms are small enough that a sponsor can decline on statistical grounds while appearing reasonable.
Still open
  1. Is there any clinical readout in this disease that is specifically a menin readout, as opposed to a general prognosis readout? If not, the class assumption may not be clinically testable short of a randomised partner-stratified menin trial.
  2. If response tracks developmental state rather than partner, what is the right stratifier for a menin trial, and can it be measured at enrolment?
  3. Does the graph's own reliance on MLL-AF9 mechanistic evidence bias it toward expecting partner equivalence, since MLLT3 is the partner with the least distinctive clinical profile?
Every source behind this entry (5)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    Partner-stratified outcome separates strongly in childhood AML: a fusion-partner-based high-risk group (n=402) had 5-year EFS 30.3% versus 54.0% (n=728), with the partner group independent of end-of-induction MRD in multivariable analysis (EFS HR 2.0, 95% CI 1.5-2.5, P<.0001).

    PMID 36996387DOI
    What this evidence cannot show

    Retrospective across protocols and eras; partner reported as a grouped binary variable; no adjustment for RAS-pathway comutation, which adult data suggest carries much of the prognostic weight; 126 patients with unassignable partners excluded.

    Evidence that argues against this (1)
    • In 205 adults with KMT2A-r AML, OS was similar across translocations while DNMT3A and KRAS mutations were independently adverse, which is the competing model in which the paediatric partner effect is unadjusted comutation. Hernández-Sánchez A et al. 2024
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper

    Where partner separates outcome, the direction depends on lineage: KMT2A::MLLT1 is in the childhood AML high-risk group and is the best-outcome partner-by-lineage cell in non-infant childhood ALL (t(11;19) T-ALL 5-year EFS 91.2% +/-4.9, n=34, versus whole-cohort 69.1%).

    PMID 36256911DOI
    What this evidence cannot show

    Cross-study juxtaposition; no partner-by-lineage interaction has ever been fitted. The favourable cell has n=34 and the contrasting adverse cell n=9. Full text not open access.

    Evidence that argues against this (1)
    • A 2,080-patient study that varied the 5-prime partner while holding MLLT10 fixed found no outcome difference at all, including for fusions containing no KMT2A, so fusion-component identity does not reliably translate into outcome. Abla O et al. 2024
  3. E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper

    Partner does not separate outcome in infants: a 175-infant BCP-ALL cohort found KMT2A::AFF1 2-year EFS 36% (95% CI 21-59) versus 37% (23-60) for other partners (P=0.9), and CCG 1953, which genotyped all 115 enrolled infants molecularly, concluded that translocation partners do not predict outcome.

    PMID 40000924DOI
    What this evidence cannot show

    Both infant cohorts are small with wide confidence intervals and the larger pools all non-AFF1 partners into one comparator arm; these are unpowered nulls, not demonstrations of equivalence.

    Evidence that argues against this (1)
    • In 465 KMT2A-r B-ALL cases across the full age range, fusion partner remained an independent determinant of MRD clearance after adjustment for age and developmental maturity (P=0.03), with AFF1 associated with slower clearance (P=7.0E-04), so a partner effect on early response exists even where the survival comparison in infants is flat. Hartmann AM et al. 2026
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    Partner does not separate outcome in adult AML: in 172 adults, multivariable analysis found translocation subtype non-predictive and the authors concluded that all KMT2A-r should be classed as adverse regardless of partner.

    PMID 34588432DOI
    What this evidence cannot show

    Single-institution retrospective series with per-arm n unreported; an unpowered null. Referral-centre case mix.

    Evidence that argues against this (1)
    • Two smaller adult cohorts do find a partner effect, with KMT2A::ELL showing 3-year OS 59.8% versus 39.3% (P=0.023), so the adult null is not unanimous. Shen H et al. 2026
  5. E5established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    No menin-inhibitor trial reports response by 3-prime partner in the primary literature (five counted PubMed queries, zero clinical reports), and AUGMENT-101's efficacy population contains 2 patients with t(4;11)/KMT2A::AFF1 and 13 with fusion partner unknown out of 57.

    PMID 39121437DOI
    What this evidence cannot show

    The AUGMENT-101 partner response numbers exist in Data Supplement Table S3, which is not carried in the PMC record and was not read; this claim is about the primary literature as retrievable, not about what sponsors hold.

    Evidence that argues against this (1)
    • The AUGMENT-101 authors state that an exploratory analysis demonstrated efficacy across translocation partners; if that supplement is adequate, the class assumption already has direct clinical support and this hypothesis loses its main motivating gap. Issa GC et al. 2025
Connected to