Analysis (runnable): build the first correlation matrix of menin-inhibitor resistance routes, from screens and from patient cohorts
Collect every published case and every screen hit where a leukemia escaped a menin inhibitor, record which escape route it used, and count how often each pair of routes occurs together. That table has never been made.
The same thing, for a scientist
A co-occurrence and mutual-exclusivity analysis across three data tiers: published CRISPR and base-editor resistance screens, published patient cohorts with resistance-mechanism annotation, and model systems with paired genotype and transcriptome. Output is a route-by-route matrix with exact-test or permutation-based significance, stratified by age group and by line of therapy. The honest expectation is that most cells of the matrix will be empty or uninformative; the deliverable is then the map of what is missing, which is the argument for what the paediatric menin trials should bank.
Route definitions, fixed before any counting
Ambiguous route definitions are how a co-occurrence analysis becomes unfalsifiable. Pre-specify:
- MEN1 interface mutation — a somatic non-synonymous variant at a residue contacting the inhibitor 1.
- PRC1.1 loss — inactivating PCGF1 or BCOR alteration, or loss of the corresponding expression, with the reduced-monocytic-signature phenotype 2.
- Signalling co-mutation — RAS pathway (KRAS, NRAS, NF1, PTPN11), FLT3-ITD, WT1 3.
- Non-genetic / cell-state escape — resistance with documented sustained on-target transcriptional effect and no route 1-3 lesion 4.
- Lineage switch — immunophenotypic conversion with clonal identity retained.
A case may be assigned to more than one route; that is the entire point. Cases that cannot be assigned are counted and reported, not dropped silently.
The three tiers and what each can support
Tier A, screens. Re-extract hit lists from the published menin-resistance CRISPR and base-editor screens and ask whether route-defining genes are recovered together or separately. Weak, because screen design determines what can be recovered at all, and a screen that only edits MEN1 cannot report PRC1.1. State that limitation rather than working around it.
Tier B, patient cohorts. Extract per-patient resistance annotation from published series, principally the post-menin-failure outcomes cohort of 84 adults 3 and any paediatric series. Count pairwise co-occurrence against the product of marginals, with Fisher exact tests and a clear statement of how many patients had all routes assessed — which will be very few, because nobody assesses non-genetic escape routinely.
Tier C, models. Paired genotype and transcriptome in resistant derivatives, where both genetic and non-genetic routes are measurable in the same sample. The most informative tier and the smallest.
The likely outcome, said in advance
This analysis will probably return an almost-empty matrix with one or two populated cells and confidence intervals spanning independence. That is not a failed analysis. The primary deliverable is the explicit accounting of how many patients in the world have had two resistance routes assessed simultaneously, which we expect to be a small number and which is the concrete argument for a standing requirement that menin trials bank paired DNA and RNA at progression.
Writing the expected outcome down before running it is the only protection against reporting a chance-significant cell as a discovery.
| script | analyses/resistance-route-cooccurrence/run.py |
|---|---|
| status | planned |
| result summary | Not yet run. No result should be quoted from this node. |
| cost tier | low |
| compute | laptop-scale |
| timeline | 3-4 weeks, most of it spent on extracting per-patient annotation from supplementary tables |
| prespecified expectation | The matrix is expected to be mostly empty with intervals spanning independence. The assessment-completeness table, not the matrix, is expected to be the useful deliverable. |
| multiple testing note | With five routes there are ten pairwise tests. Report all ten with adjusted values; reporting one nominally significant cell without the other nine would be the obvious abuse. |
inputs
[ "Published menin-resistance CRISPR and base-editor screen hit lists (PMID 36922589, PMID 39158067)", "Per-patient resistance annotation from post-menin-failure cohorts (PMID 41351513) and any paediatric series", "Model-system paired genotype and transcriptome from resistant derivatives (PMID 41027490, PMID 38892207)", "Project EVOLVE lineage-switch cases with molecular annotation (PMID 40193715)" ]
outputs
[ "route-by-route co-occurrence matrix with observed versus expected counts and exact-test or permutation p-values", "an explicit assessment-completeness table: how many patients had each pair of routes actually assessed", "stratification by age group and line of therapy where n permits", "a short specification of what menin trials would need to bank to fill the empty cells" ]
tests hypotheses
[ "hyp-resistance-routes-negatively-correlated-genetic-vs-plastic" ]
- An almost-empty matrix invites over-interpretation of the few populated cells. The pre-specified expectation and the full ten-test reporting rule exist to stop that.
- Are per-patient resistance annotations available in supplementary tables, or only as aggregate percentages, in which case Tier B collapses to nothing?
- Is there an existing consortium already assembling menin-failure samples that this analysis should feed into rather than duplicate?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
MEN1 drug-interface mutations were identified in patients with acquired resistance to menin inhibition and confirmed in xenografts and an unbiased base-editor screen, giving a precise residue-level definition for route 1 and a screen dataset for Tier A.
MEN1 mutations mediate clinical resistance to menin inhibition · Perner F, Stein EM, Wenge DV, et al. · Nature 2023What this evidence cannot show
The base-editor screen targeted MEN1 and therefore cannot report on other routes; this is the structural limitation of Tier A.
Evidence that argues against this (1)
- One assumption behind treating MEN1 interface mutation as a cleanly separable route needs stating: the mutations attenuate small-molecule binding while preserving the menin-MLL1 interaction, and they arise readily de novo in an unbiased base-editor screen. A route that is generated on demand under selection, rather than drawn from standing variation, will show frequency patterns driven by exposure and follow-up rather than by competition with other routes — which is a confounder for any co-occurrence inference, and one that regimen-matching cannot fully remove. Perner F et al. 2023
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Chromatin-focused CRISPR screens identified PRC1.1 subunit loss (PCGF1, BCOR) as a menin-inhibitor resistance route reactivating non-canonical menin targets including MYC, with a reduced monocytic signature and BCL-2 dependence as the associated phenotype.
Epigenetic regulation of noncanonical menin targets modulates menin inhibitor response in acute myeloid leukemia · Zhou X, Zhang L, Aryal S, et al. · Blood 2024What this evidence cannot show
Screen hit lists provide the gene set; patient frequency for this route is zero-reported, so Tier B will have an empty column for it.
Evidence that argues against this (1)
- THE PLANNED OUTPUT IS LIKELY TO BE NOISE. This analysis exists to estimate the off-diagonal terms of a resistance-route covariance matrix and use them to rank combinations. The portfolio literature that supplies the frame has already tested whether estimated covariance matrices support allocation decisions, and the answer is no: across seven empirical datasets, none of fourteen mean-variance and extended optimisation models consistently beat the naive 1/N rule on Sharpe ratio, certainty-equivalent return or turnover, because out of sample the gain from optimal diversification is more than offset by estimation error. That is with decades of daily data on liquid assets. Here the inputs are a handful of CRISPR screens with different libraries, cell contexts, selection pressures and hit-calling thresholds. The analysis should be scoped to the qualitative question — are any two routes significantly non-independent, in which direction — with explicit multiple-testing control, and should not produce a combination ranking. DeMiguel V et al. 2009
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
In 84 adults with relapsed or refractory AML after menin-inhibitor failure, FLT3-ITD, WT1 and MEN1 mutations were associated with resistance, with per-patient mutation annotation available, giving the principal Tier B cohort.
Outcomes of relapsed or refractory acute myeloid leukemia after menin inhibition failure · Multicentre retrospective consortium · Blood Advances 2026What this evidence cannot show
Adults only, retrospective, and non-genetic escape was not assessed, so co-occurrence between genetic and non-genetic routes cannot be computed from this cohort.
Evidence that argues against this (1)
- Searched for any published dataset in which the named menin-inhibitor resistance routes — MEN1 interface mutation, PRC1.1 loss, cell-state/transcriptional escape and signalling co-mutation — were scored jointly in the same patients or the same screen, which is the minimum requirement for estimating co-occurrence. None was found. The per-patient mutation annotation in the post-failure cohort covers genetic routes only (FLT3-ITD, WT1, MEN1) and carries no chromatin or cell-state scoring, so even the best available dataset supports at most the genetic sub-block of the matrix. The analysis should say this explicitly and report which cells of the matrix are estimable at all. Recorded so the gap is known to have been checked rather than skipped. Q12-analogy-counter (search record) 2026
- E4moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
Menin-inhibitor resistance occurred in pretreated KMT2A::AFF1 samples with sustained on-target transcriptional effect and without prior drug exposure, providing the operational definition of non-genetic escape used here.
Upfront menin-inhibitor resistance in multiply pretreated leukemias · Mahdavi L, Bernt KM, et al. · Experimental Hematology 2025What this evidence cannot show
Two patients; adequate to define the route operationally and inadequate to contribute to a frequency estimate.
Evidence that argues against this (1)
- This defines non-genetic escape by exclusion — on-target effect sustained, no prior drug exposure — and the same samples carried emergent RAS and TP53 lesions. An operational definition that rests on not having found the genetic explanation will classify undiscovered genetics as plasticity. Nature 2023
estimation-error objection: across seven datasets no optimisation model beat naive 1/N because covariance estimation error dominates (DeMiguel 2009), so the planned combination ranking should be dropped in favour of a qualitative substitutes/complements test
builds the route-by-route co-occurrence matrix and the assessment-completeness table
adds a downstream/apoptotic route to the route-by-route co-occurrence matrix
supplies four additional named routes for the route-by-route matrix