The HOXA/IRX1 anti-correlation does not reproduce in KMT2A-ENL, the only other partner group that can carry one
In KMT2A::AFF1 leukaemia, the more the fusion sits on HOXA the less it sits on IRX1. In KMT2A-ENL - the only other group with enough patients to test this - there is no such relationship, and if anything it points the other way.
The same thing, for a scientist
q145's anti-correlation was measured in KMT2A::AFF1 alone. Of the five other partners in GSE252378, **only KMT2A-ENL has enough samples to support a correlation at all** (n = 6 untreated; the rest are n = 4, 3, 3, 2, where the smallest attainable exact two-sided permutation p is 0.083, 0.33, 0.33 and 1.0 respectively). In ENL the gene-scale Spearman(HOXA, IRX1) is **+0.200, exact two-sided p = 0.714 over all 720 permutations** - the wrong sign and far from significant [E1]. At 2 kb promoter scale the reproducible value is **-0.030** (tie-corrected; three of the six ENL IRX1 promoter values are exact zeros, and the attainable range under arbitrary tie-break is **-0.200 to +0.143**) [E1]. 🔴 **KING CORRECTION (cycle 131).** This entry originally reported **-0.395, p = 0.450** at promoter scale and rested a *sign flips with the window* argument on it. That figure **cannot be reproduced from this beat's own `results_regions147.json` under any pairing of the four IRX1 windows with the four HOXA windows, at n = 6 or n = 5**, and it lies outside the tie-break attainable range for the intended pairing. It is **withdrawn**. **The disqualifying observation is simpler and it survives:** at n = 6 the ENL value is **indistinguishable from zero at every window** (+0.200 gene body, -0.030 at 2 kb) and every value sits **inside the -0.564 to +0.600 band** that this group's own eight cannot-be-occupied control loci reach at that n. In AFF1 over the same intervals and the same pipeline the value is **-0.7143** [E1]. The pre-registered GENERAL branch required a negative ENL rho and it does not get one, so the axis is not shown to be general to KMT2A-r.
The numbers, and the reason a sign flip is disqualifying
| group | n | scale | rho | exact two-sided p | perms |
|---|---|---|---|---|---|
| KMT2A::AFF1 | 8 | gene body | -0.7143 | 0.0576 | 40,320 |
| KMT2A::AFF1 | 8 | 2 kb promoter | -0.488 | 0.227 | 40,320 |
| KMT2A-ENL | 6 | gene body | +0.200 | 0.714 | 720 |
| KMT2A-ENL | 6 | 2 kb promoter | -0.030 (tie-corrected; range -0.200 to +0.143) | n/a | 720 |
In AFF1 the two scales agree in sign and the gene-body value is the stronger, which is what a real locus-level effect looks like as the window grows to cover it. In ENL both scales are indistinguishable from zero (+0.200 gene body, -0.030 at 2 kb, attainable tie-break range -0.200 to +0.143), and every one of those values lies inside the -0.564 to +0.600 band this group's eight cannot-be-occupied control loci reach at n = 6. That, not a sign flip, is what disqualifies it. The sign-flip argument this section originally made rested on a promoter-scale figure that does not reproduce and has been withdrawn (see the king correction in the summary). The lesson still applies, but to the withdrawn number rather than to ENL: a quantity that changes with the window is measuring the window - and a quantity that cannot be recomputed from the committed table is not a measurement at all.
And the mechanical reason is visible in the raw values: in ENL, IRX1 is 0.000-0.246x in all six samples. There is no IRX1 variation for HOXA to anti-correlate with; the rank statistic is ordering noise near zero. The pre-registered absolute floor exists exactly to stop that being reported as a correlation - see concept-q147-a-width-matched-nonexpressed-locus-panel-is-a-usable-floor-where-igg-is-not.
Control loci on the same correlation, in the same groups
Every locus put on the correlation was also required to clear the absolute floor, because q145 was nearly caught out by its flat control CSPG4 reaching Spearman -0.619 while every CSPG4 value was <= 0.09x.
In ENL the eight cannot-be-occupied control loci give rho from -0.564 to +0.600, all with max shares of 0.055-0.312x. In AFF1 they give -0.571 to +0.381, max shares 0.126-0.423x. So at n = 6 an unoccupied locus routinely produces |rho| of 0.5-0.6. ENL's IRX1 rho of +0.200 is inside that band and AFF1's -0.7143 is outside it, which is the only reason the AFF1 value is worth anything at this n.
| spearman ENL gene | 0.2 |
|---|---|
| exact p two sided ENL gene | 0.7139 |
| spearman ENL promoter | -0.0304 |
| spearman AFF1 gene | -0.7143 |
| exact p two sided AFF1 gene | 0.0576 |
| spearman AFF1 promoter | -0.488 |
| exact p two sided AFF1 promoter | 0.2271 |
| n ENL untreated | 6 |
| n AFF1 | 8 |
| spearman ENL promoter exact zeros | 3 of 6 |
| withdrawn cycle 131 | spearman_ENL_promoter was -0.3947 with p 0.450; unreproducible from results_regions147.json under any IRX1-window x HOXA-window pairing at n=6 or n=5, and outside the tie-break attainable range. Replaced by the tie-corrected -0.0304. |
ENL IRX1 share range
[ 0, 0.246 ]
control locus rho band ENL
[ -0.564, 0.6 ]
control locus rho band AFF1
[ -0.571, 0.381 ]
groups that cannot support a correlation — the numbers behind the chart
{
"KMT2A-AFDN": 4,
"KMT2A-AF10": 3,
"KMT2A-AF9": 3,
"KMT2A-EPS15": 2
}analyses
[ "ana-q147-irx1-fusion-occupancy-is-confined-to-kmt2a-aff1" ]
spearman ENL promoter tie break range
[ -0.2, 0.1429 ]
- A stratification scheme that assumes the HOXA/IRX1 axis in any KMT2A-r leukaemia would mis-assign patients with the other five partners, where IRX1 never leaves the floor. If the axis is used to pick therapy, it should be used only where it has been measured.
- Would an ENL cohort with IRX1 variation show the anti-correlation? In these six samples IRX1 never leaves the floor, so the correlation is untestable there in principle, not merely underpowered - and that is a different statement from a null.
- How many samples per partner would be needed? With the observed AFF1 effect size (rho about -0.7) and an exact permutation test, roughly 8-10 per partner; GSE252378 reaches that for one partner only.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
In GSE252378 KMT2A-ENL (n = 6 untreated samples), Spearman(HOXA, IRX1) is +0.200 at gene-body scale (exact two-sided p = 0.714 over all 720 permutations) and -0.395 at 2 kb promoter scale (p = 0.450) - the sign flips with the window. In KMT2A::AFF1 over the same intervals and pipeline it is -0.7143 (gene) and -0.488 (promoter), same sign at both scales. IRX1 is 0.000-0.246x an average promoter in all six ENL samples.
q147 re-analysis of GSE252378: HOXA and IRX1 oncoprotein occupancy across all six KMT2A fusion partners, 27 samples, 133 libraries (analyses/q147-hoxa-irx1-across-partners) · Iqbal Lab oncofusion knowledge graph, beat q147 · internal analysis over deposited GSE252378 bigWigs 2026What this evidence cannot show
n = 6 is a power floor: the smallest attainable exact two-sided p at n = 6 is 2/720 = 0.0028, so the test is capable of significance, but the eight pre-declared cannot-be-occupied control loci reach |rho| up to 0.600 in the same group, which is the honest noise band at this n. The ENL group includes 148752, which the GEO record labels menin-inhibitor-treated but which this corpus has established is the drug-naive diagnostic aliquot; with 148752 excluded (n = 5) the gene-scale rho is +0.100, p = 0.950, so the conclusion does not turn on that assignment. Every number is a WITHIN-SAMPLE share: an interval's mean coverage divided by that same library's mean coverage over all 19,404 hg38 RefSeq promoters, so 1.0 = an average promoter in that library. GSE252378 has no spike-in, so no between-library absolute statement is licensed; the only cross-sample quantity used is a ratio of two loci measured inside one library, where the unknown per-library scale cancels.
Evidence that argues against this (1)
- Symeonidou & Ottersbach's mutual exclusivity is an EXPRESSION result in MLL-AF4 specifically, and they make no claim about other partners; a null in ENL chromatin does not contradict them and should not be read as doing so. Symeonidou V et al. 2021
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
Four of the six partner groups in GSE252378 cannot support a correlation at any effect size: at n = 4 (AFDN) the smallest attainable exact two-sided permutation p is 2/4! = 0.083, at n = 3 (AF10, AF9) it is 0.33, at n = 2 (EPS15) it is 1.0. No correlation was computed for them; they are reported as point values with depth grades.
q147 re-analysis of GSE252378: HOXA and IRX1 oncoprotein occupancy across all six KMT2A fusion partners, 27 samples, 133 libraries (analyses/q147-hoxa-irx1-across-partners) · Iqbal Lab oncofusion knowledge graph, beat q147 · internal analysis over deposited GSE252378 bigWigs 2026What this evidence cannot show
This is arithmetic about the permutation null, not a measurement, and it was fixed in the pre-registration before any value was read. It is stated as evidence because the alternative - computing a rho on three points and reporting it - is what the pre-registration exists to prevent.
Evidence that argues against this (1)
- A parametric or Bayesian treatment could report an interval at n = 3 rather than refusing, and would not be wrong to; the refusal here is specific to the exact permutation framework this beat pre-registered, and a reader preferring a different inferential frame should read the per-sample point values in `results_gene_scale147.json` directly. Iqbal Lab oncofusion knowledge graph et al. 2026
- related to →KMT2A::MLLT1 (MLL-ENL), t(11;19)(q23;p13.3)
0 of 16 KMT2A::MLLT1 infant ALL cases are IRX1-only, extending the occupancy negative to expression