The GMP-like peak class is fusion-enriched in exactly one of GSE252378's 27 KMT2A-rearranged samples - p318, the sample it was derived from - once the null is matched on how coverable those promoters generically are
The published idea is that when a leukaemia changes identity, the KMT2A fusion protein moves onto the genes of a myeloid progenitor. Scored at the gene switches of all 27 patients in the deposit, that myeloid gene set looks busy everywhere - but so does any equally readable set of genes, and once that is corrected for, only the single patient the gene set was built from still stands out.
The same thing, for a scientist
The authors' own `AML_specific` peak class - their export name for it is *Supplementary_Information_Table_GMP-like_Genes.csv* - resolves to **40 promoters** in the 19,404-promoter hg38 set. Scored as a within-library share over the KMT2A-N (fusion-side) libraries of all 27 GSE252378 samples it spans **1.310x-3.739x an average promoter in the same library**, and across the 23 untreated non-circular samples **1.305x-2.607x** - a tight band that clears a naive uniform permutation null in **18 of those 23**. **That is an artefact of the null.** The class is a set of generically high-coverage promoters: against a null matched on coverability decile (the median share across the KMT2A-N libraries of the *other* 26 samples), it is enriched in **0 of 23**, and in **1 of all 27** - **p318**, the lineage-switched infant AML in which the class was defined, at **1.767x its matched null** (two-sided permutation p = 0.0004 and 0.0052 in its two N libraries). The same matched null leaves the canonical `Oncoprotein_Target` class enriched in **16 of 23** and the `ALL_specific` class enriched in **7 of 23**, so the instrument is not blind - it is specifically the GMP-like class that does not survive.
The two nulls, side by side, with the denominator
| class | n promoters | calls OCCUPIED under the uniform null N2 | enriched over a coverability-matched null N4 |
|---|---|---|---|
Oncoprotein_Target | 827 | 21 of 23 | 16 of 23 |
ALL_specific | 37 | 3 of 23 | 7 of 23 |
AML_specific (GMP-like) | 40 | 18 of 23 | 0 of 23 |
Denominator: the 23 samples that are untreated, not one of the two that defined the classes, of GSE252378's 27. N2 draws random promoter sets of the same size uniformly from the 19,404. N4 draws them matched on the decile of the median share across the KMT2A-N libraries of the other 26 samples, so it asks is this class above an equally coverable promoter rather than above an average promoter. N2 and N4 disagree on 18 of 23 samples for this one class, and N4 is the right question to ask of a curated gene set.
The one sample where it does survive, and why that is not a positive result
p318 - infant KMT2A::AFF1, Texas Children's, the post-switch AML half of the p279/p318 pair - carries AML_specific at 3.739x against a matched null of 2.12/2.11, ratio 1.767, the only UP call in the deposit. It is also the sample the class was defined in: the authors' construction takes intervals with z < -1.96 in a p279-minus-p318 comparison and restricts them to genes absent from the canonical oncoprotein-target list. Scoring a class in the sample that defined it is circular, and it is recorded here as the reagent's positive control, not as evidence.
What that leaves is the honest bound: GSE252378 contains two post-lineage-switch AML samples, p318 and 152985. p318 is circular for this class and 152985 is the revumenib-exposed aliquot in which every channel collapses (AML_specific 1.310x, matched-null ratio 0.976, ns). So the number of independent, non-circular, drug-naive post-switch samples available to test the published GMP claim in this deposit is zero. The third lineage-switch pair, R3595/R3652 (KMT2A-EPS15, Seattle Children's), is labelled B-ALL on both halves, so it contributes no post-switch AML sample at all; its AML_specific matched-null ratios are 0.958 and 1.059, both ns.
What this does and does not say about the published claim
The paper's abstract states that "in lineage-switching samples the oncoprotein levels are reduced and the oncoproteins preferentially activate granulocyte-monocyte progenitor (GMP) genes" 1. This analysis does not refute that. It measures a different window (2 kb promoters, not the 5-20 kb peak intervals the authors used), on 40 of their genes, with a different normalisation, and it agrees with them wherever they had signal: p318 is the one sample with a GMP-like excess and it is a lineage-switching sample.
What it adds is a bound. The GMP-like programme is not a general destination for the KMT2A oncoprotein: it is not elevated over an equally coverable promoter set in any of the 23 other samples, across six fusion partners, four institutions and three lineages. And in particular it is not where the oncoprotein goes in the one depth-adequate HOXA-negative, IRX1-negative blast, A73051.
The reagent's losses, stated because a filter must report what it removed
From 2,685 CSV rows the authors' file resolves to 1,692 unique intervals - 52 ALL_specific, 45 AML_specific, 1,595 Oncoprotein_Target, exactly the n = 52 and n = 45 that the q143 beat scored at peak scale. Converting intervals to promoters loses:
- 284 intervals with no annotated gene - invisible to a promoter-window reagent;
- 29 genes annotated to more than one class, dropped from all three by the pre-registered exclusivity rule (24 of them from
ALL_specific, which is why that class falls from 66 raw genes to 42); - 300 genes that do not join the 19,404 promoter symbols (5
ALL_specific, 6AML_specific, 289Oncoprotein_Target), overwhelmingly lncRNA, antisense andLOCsymbols that a RefSeq canonical-TSS set does not carry.
Final sizes: 37, 40 and 827 promoters.
| deposit | GSE252378 |
|---|---|
| n samples | 27 |
| n libraries | 133 |
| n promoters | 19404 |
| units | within-sample promoter share; 1.0 = an average promoter in the SAME library |
| no spike in | true |
| script | analyses/q148-what-the-hoxa-negative-blast-occupies/analyze148.py |
| status | run |
| result summary | GMP-like class enriched over a coverability-matched null in 0 of 23 untreated non-circular samples, 1 of 27 overall, and that one is the sample the class was defined in. |
geneset sizes — the numbers behind the chart
{
"ALL_specific": 37,
"AML_specific": 40,
"Oncoprotein_Target": 827
}reagent losses
{
"intervals_with_no_gene": 284,
"ambiguous_genes_dropped": 29,
"genes_failing_promoter_join": {
"ALL_specific": 5,
"AML_specific": 6,
"Oncoprotein_Target": 289
}
}measurements — the numbers behind the chart
[
{
"quantity": "n_samples_with_GMP_class_enriched_over_matched_null",
"subject": "AML_specific (GMP-like) 40-promoter class, KMT2A-N libraries",
"value": 0,
"units": "samples",
"n": 23,
"universe": "GSE252378 untreated, non-circular (p279/p318 excluded) samples, 19,404-promoter hg38 set",
"null": "coverability-decile-matched permutation, 10,000 draws, leave-the-sample-out",
"note": "1 of all 27 counting p318, which defined the class"
},
{
"quantity": "n_samples_with_GMP_class_called_occupied_by_the_uniform_null",
"subject": "same class, same libraries",
"value": 18,
"units": "samples",
"n": 23,
"universe": "same 23 samples",
"null": "uniform permutation over the 19,404 promoters",
"note": "the pre-registered N2 gate; it disagrees with the matched null on 18 of 23"
},
{
"quantity": "observed_over_matched_null",
"subject": "p318 AML_specific class",
"value": 1.767,
"units": "ratio",
"n": 2,
"universe": "p318 KMT2A_N1 and N2, 40 promoters",
"null": 1,
"note": "two-sided permutation p 0.0004 and 0.0052; CIRCULAR - p318 defined the class"
},
{
"quantity": "n_samples_with_canonical_class_enriched_over_matched_null",
"subject": "Oncoprotein_Target 827-promoter class",
"value": 16,
"units": "samples",
"n": 23,
"universe": "same 23 samples",
"null": "coverability-matched permutation",
"note": "the instrument is not blind; the GMP class specifically fails"
}
]- If the GMP-like state is being treated as a therapeutic anchor - the state a menin inhibitor is said to hit hardest - then an occupancy measurement that finds no GMP-like oncoprotein enrichment outside one patient should make anyone pause before selecting patients on that state.
- A negative that rests on a matched null can be an over-correction. If the fusion genuinely occupies generically accessible promoters, matching on accessibility removes real signal, and a therapy decision taken on this negative would be taken on a conservative instrument.
- Does the GMP-like class survive a coverability-matched null when scored at the AUTHORS' peak intervals rather than at 2 kb promoters? The bigWigs are public and the interval list is on disk; this is the single cheapest way to separate my window from their finding.
- Is there any deposited post-lineage-switch, drug-naive, non-p318 KMT2A-rearranged sample anywhere with paired N and C terminal occupancy? GSE252378 has none, and that is the gap that makes the published GMP claim untestable outside its founding pair.
- Would a GMP gene set defined from normal haematopoiesis rather than from one patient pair give the same answer? The class used here is 40 promoters chosen by a difference between two samples, and a reference GMP signature is a different and less circular reagent.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The deposit's own paper reports that in lineage-switching samples the oncoprotein levels are reduced and the oncoproteins preferentially activate GMP genes.
MLL oncoprotein levels influence leukemia lineage identities. · Janssens DH, Duran M, Otto DJ, Wu W, Xu Y, Kirkey D, Mullighan CG, Yi JS, Meshinchi S, Sarthy JF, Ahmad K, Henikoff S · Nature Communications 15:9341 2024What this evidence cannot show
Read from the abstract and the figure-generation notebook, not from a re-reading of every figure. The claim in the paper rests on three lineage-switching samples out of 36 profiled, and GSE252378 deposits only two post-switch AML samples, one of which is menin-inhibitor exposed.
Evidence that argues against this (1)
- The authors' construction defines the GMP-like class from the p279/p318 comparison and then reports GMP-gene activation in lineage-switching samples, so p318 is inside both the definition and the claim. That is not an error in their paper - the class is offered as a description of that pair - but it does mean a reader must not count p318 as independent confirmation. Janssens DH 2024
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Scored at 2 kb promoters against a coverability-matched null, the AML_specific (GMP-like) class is enriched in 0 of 23 untreated non-circular GSE252378 samples and in 1 of 27 overall - p318, at 1.767x its matched null.
analyses/q148-what-the-hoxa-negative-blast-occupies - results148.json, results148_matched_null.json; 133 libraries x 19,404 promoters, pre-registered 2026-09-24 · Q148 (Oncofusion Knowledge Graph) · this repository 2026What this evidence cannot show
Promoter windows are not the peak intervals the class was built on, and 40 promoters is a small set. The matched null is POST-HOC - the pre-registration specified an IgG-decile-matched null, which could not be computed because IgG promoter shares are 9%-98% zero across the 25 samples that have an IgG library. Both nulls are reported; neither is presented as replacing the other.
Evidence that argues against this (1)
- A coverability-matched null can be too conservative: if the fusion genuinely occupies generically accessible promoters, matching on coverability partly matches on the signal. The control against that reading is that the same matched null still returns 16 of 23 enriched for the canonical Oncoprotein_Target class and 7 of 23 for ALL_specific, so it is not simply subtracting all signal. Q148 (Oncofusion Knowledge Graph) 2026
- E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
GSE252378 deposits 171 samples, all with !Sample_library_strategy = OTHER and no mention of RNA-seq, so every quantity here is chromatin occupancy and none is expression.
GSE252378, Gene Expression Omnibus - KMT2A oncoproteins induce epigenetic resistance to targeted therapies; 171 sample records, Series_pubmed_id 39472576, read from the SOFT family file 2026-09-24 · Janssens DH, Ahmad K, Henikoff S · NCBI Gene Expression Omnibus 2024What this evidence cannot show
Counted on the locally cached SOFT family file: 171 of 171 GSMs carry library_strategy OTHER and the file contains zero case-insensitive matches for 'RNA-seq' or 'transcriptome'. That bounds the deposit, not the literature.
Evidence that argues against this (1)
- The published HOXA9/IRX1 subgroup split that motivates this whole line of work is an EXPRESSION result from two RNA-seq cohorts with no chromatin experiment, so an occupancy negative here cannot refute it and an occupancy positive cannot confirm it. Symeonidou V et al. 2021
extends q143's two-sample peak-scale scoring of the same classes to all 27 samples at promoter scale
the GMP-like chromatin class is not fusion-enriched outside the pair that defined it
answers that entry's own open question: is a GMP gene set fusion-occupied in TB920488 and A73051
scores the chromatin-wise GMP-like signature across six partners; enriched nowhere but p318
- depends on →p279 and p318: one infant, two diseases, and an oncoprotein target map that does not carry over
p279/p318 is the pair the peak classes were derived from
- related to →Granulocyte-monocyte progenitor (GMP) — the committed cell an MLL fusion can turn back into a stem cell
the class is the authors' GMP-like gene export, scored as occupancy not expression
the branch outcome of the measurement in that entry
a second, normalisation-independent route to the same conclusion
explains why a stable 1.3-2.6x numerator is not occupancy
the reagent audit for that measurement's B-lineage class
the reagent calibration that measurement rests on
the methodological reason that measurement's verdict changed
bounds that measurement to occupancy language