Verdict: q147's IRX1 headline survives as a measurement and not as an inference - the pre-registered branch is NEITHER, on U5 and on a FALLS condition that fired
also called q154 verdict · does the IRX1 address survive its null
The question was whether the fusion protein sitting on IRX1 in three AFF1 patients and none of eighteen others is more than chance. Measured against a proper null at the right scale, it is: only six genes in the whole genome do that, where chance gives about one. But those three samples are two children from one hospital, and once you count them as two patients the result stops being significant - so the finding is real as a measurement and cannot yet carry the conclusion that IRX1 needs the AFF1 partner.
The same thing, for a scientist
**Branch: NEITHER.** Scored against the q154 pre-registration in the order it specifies. **SURVIVES needed all four.** (1) The class above its null under both floors: **met** - 6 against 0.86 exact institution-stratified, p_upper 0.0011, identical membership under three floors. (2) T3 patient-level p <= 0.05: **NOT met** - 0.0772. (3) The class smaller in KMT2A-C than in KMT2A-N: **met** - 0 against 6. (4) Leave-one-AFF1-out above the null: **met** - 4 to 6 across all eight drops. **Three of four.** **FALLS needed any one.** (1) The class at or below its null: **no**. (2) T3 patient-level p > 0.20: **no**, 0.0772. (3) The class as large in KMT2A-C: **no**. (4) **IgG within a factor of two of KMT2A-N: FIRED** - IgG passes a one-library form of the gate at 1.938% against KMT2A-N1's 0.948%, a factor of 2.0 in the unexpected direction. **That condition is overridden with a measurement, not an argument**: IgG's promoter density is 8.45-fold lower than KMT2A-N1's, so the within-sample-share denominator collapses in an IgG library and the control's unit is defined by the enrichment it is supposed to lack. **The condition fired, it is recorded as having fired, and the control is invalid for a stated measured reason.** **U5 also fires as written**: the verdict differs between sample level (p 0.0026-0.0011) and patient level (0.1164) for the class, and between sample level (0.0204) and patient level (0.0772) for IRX1 itself. **U5 alone forces NEITHER.** U1a, U1b, U2, U3, U4 and U6 do not fire: the instrument gate passes at Spearman 0.9901 and reproduces q147's call table exactly, the reference class is 148 not under 30, the floor at IRX1's length decile exceeds 2.0 in 3 of 54 libraries not in over half, the run completed 133 of 133 with 0 failures, and no observed value is within 2 Monte-Carlo standard errors of its null.
The plain answer the brief asked for
Is 0-of-18 surprising against a gene-scale null, or is it what chance gives?
0-of-18 on its own is what chance gives. At a 0.608% gene-body pass rate, an arbitrary gene body is expected occupied in 0.164 of 27 samples, so being absent from 18 comparators is nearly free: 63 gene bodies are occupied in >=1 AFF1 sample and 0 of 18, against a permutation null of 73.9 - at or below chance, exactly as at promoter scale.
What is surprising is the conjunction with recurrence. Three or more of the eight AFF1 samples and zero of eighteen is done by 6 gene bodies against a null of 0.86 - about seven times chance, p 0.0011 under an exhaustive institution-stratified permutation, and stable across three floor definitions and eight leave-one-out drops.
And the conjunction does not survive counting patients. 12 against 6.35, p 0.1164. IRX1 itself: p 0.0204 by sample, 0.0772 by patient, 0.1190 holding institution fixed, 0.1756 restricted to B-lineage.
So: q147's headline is a correct measurement of a real pattern that is not what chance gives at sample level, and it does not have the statistical support to carry the inference that the IRX1 address requires AFF1. It should not be withdrawn. It should be restated with its n: two independent patients, one institution.
What makes IRX1 stronger than the class it belongs to
The class of six is mostly a threshold artefact and IRX1 is not. Five of the six have a non-AFF1 sample within a factor of 2.2 of the 2.0x gate - H3C15 within 1.4% - and the deposit paper's own FDR-based caller calls three of them (ELF1, GNAQ, PAN3) in KMT2A-ENL and KMT2A-AF9 samples. IRX1's nearest non-AFF1 sample is at 0.252x, 7.9-fold below the gate, and the authors' caller lists IRX1's samples as exactly ['p279','p318','p454'].
Two independent instruments on the same libraries, one a fold gate and one a peak caller with a Bayesian N/C model, agree on IRX1's sample set and disagree on three of its five classmates. That is the strongest single fact this beat produced and it is not a null - it is a reproduction.
The scale disagreement, resolved
Three cycles read a contradiction between IRX1 looking strong at gene-body scale and weak at promoter scale. It is a denominator, not a contradiction. The same rule passes 10.75% of promoters and 0.608% of gene bodies - 17.7-fold apart - so 2 of 27 at promoter scale is 0.69x expectation and 3 of 27 at gene-body scale is 18.3x. Neither number was interpretable without the other's pass rate, and until this beat neither existed.
The promoter-scale dismantling of the IRX1 evidence therefore stands as a statement about the promoter-scale instrument and does not transfer to gene-body scale. Equally, the gene-body result does not rescue the promoter-scale claims: the 332-promoter AFF1-only set is still below its null and the 17-promoter core is still at chance.
| script | analyses/q154-gene-scale-occupancy-null/analyze.py |
|---|---|
| status | run |
| prereg branch fired | NEITHER |
| result summary | SURVIVES 3 of 4 (fails T3 patient-level p 0.0772); FALLS 0 of 4 substantive but the IgG condition fired and is overridden with an 8.45-fold denominator measurement; U5 fires on sample-versus-patient disagreement. q147's headline stands as a measurement, not as an inference. |
inputs
[ "analyses/q154-gene-scale-occupancy-null/PREREGISTRATION.md", "analyses/q154-gene-scale-occupancy-null/results.json", "analyses/q154-gene-scale-occupancy-null/results_authors_null.json" ]
gates that fired
[ "U5 - the verdict differs between sample level and patient level", "FALLS condition 4 (IgG) - fired and overridden with a measured reason, recorded not dropped" ]
gates that did not fire
[ "U1a instrument Spearman 0.9901 >= 0.90", "U1b IRX1 calls reproduce q147 exactly", "U2 reference class 148 >= 30", "U3 floor above 2.0 in 3 of 54 libraries at IRX1's length decile, not over half", "U4 run completed 133 of 133, 0 failures", "U6 no observed value within 2 Monte-Carlo SE of its null" ]
measurements
[
{
"quantity": "patient_level_permutation_p_for_the_recurrent_class",
"subject": "gene bodies occupied in >=2 of 7 AFF1 patients and 0 of 17 non-AFF1 patients, against a label permutation",
"value": 0.1164,
"units": "one-sided permutation p, +1 corrected",
"n": 10000,
"universe": "19,402 gene bodies x 24 independent patients; observed 12 against a null mean of 6.35",
"null": 0.5,
"note": "the sample-level equivalent is 6 against 0.86 at p 0.0011; the difference between them is what makes the branch NEITHER"
},
{
"quantity": "patient_level_permutation_p_for_irx1_alone",
"subject": "IRX1 occupied in 2 of 7 KMT2A::AFF1 patients and 0 of 17 non-AFF1 patients",
"value": 0.0772,
"units": "one-sided permutation p, +1 corrected",
"n": 10000,
"universe": "24 independent patients after collapsing p279/p318 and R3595/R3652; 152985 dropped as drug-exposed",
"null": 0.5,
"note": "0.0204 at sample level, 0.1190 holding institution fixed (exact over 4,410 assignments), 0.1756 restricted to B-lineage"
}
]- If the IRX1 address is used to stratify KMT2A::AFF1 patients for a menin-inhibitor or an IRX1-directed strategy, the evidence behind the stratifier is two children from one hospital plus a supplementary spreadsheet row. A trial powered on it would be powered on an effect whose patient-level p is 0.077.
- The five classmates are the warning: a partner-specific target list drawn from the same analysis would contain a canonical pan-KMT2A-fusion target (PAN3), a histone gene whose nearest comparator misses the cut by 1.4% (H3C15), and two genes the data's own authors call in other partners.
- One more untreated KMT2A::AFF1 patient from outside Texas Children's moves the patient-level p from 0.077 to roughly 0.02 if IRX1 is occupied in it. No second deposit exists - a GEO series search for KMT2A or MLL-AF4 with CUT&RUN or ChIP-seq returns nothing carrying per-patient oncoprotein occupancy in this fusion - so this needs one sample and one plate.
- Does the same permutation with a KMT2A-ENL or KMT2A-AFDN index arm also give about six recurrent exclusive gene bodies? If it does, partner-exclusivity as a concept has no purchase in this deposit at any scale, and that single control would settle more than anything else listed here.
- The gene-body rule and the authors' caller make comparable numbers of calls per sample - 113 against 95 - and agree on only 21% of the authors' calls. Which disagreements are gene-body-versus-promoter geometry and which are threshold?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Scored against the q154 pre-registration, SURVIVES is met on 3 of 4 conditions and fails on the patient-level p of 0.0772; FALLS is met on 0 of 4 substantive conditions and fired on the IgG control, which an 8.45-fold promoter-density difference shows to be invalid; U5 fires because the sample-level and patient-level verdicts differ; U1a, U1b, U2, U3, U4 and U6 do not fire.
Q154 genome-wide gene-body occupancy matrix, label permutations and pre-registration gates · Iqbal Lab oncofusion knowledge graph, beat Q154 · this project 2026What this evidence cannot show
The pre-registration was written and executed by the same agent in one session; only the git commit order separates them. Two of the four stratifications used here - the exact institution enumeration and the B-lineage restriction - were specified after the per-sample institution and lineage strings were read from the deposit metadata, which happened after the file was written. Both are corrections that weaken the result, so the deviation does not flatter it, but it is a deviation and is recorded as one.
Evidence that argues against this (1)
- A NEITHER verdict reached partly on a control the author then declares invalid is the shape this project has been wrong in before. The honest minimum is that the reader can check the override: the promoter-density measurement is in the results file, the pre-registered wording is in the committed pre-registration, and both are named here. Iqbal Lab oncofusion knowledge graph et al. 2026
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
q147's headline is that IRX1 is fusion-occupied in 3 of 8 KMT2A::AFF1 samples and 0 of 18 non-AFF1 across five other partners at gene-body scale, with HOXA occupied in 17 of the same 18 as the positive control; this beat reproduces that table exactly from a different instrument and does not contradict any of its numbers.
Q147 gene-scale results - IRX1 and HOXA occupancy across six KMT2A fusion partners · Iqbal Lab oncofusion knowledge graph, beat Q147 · this project 2026What this evidence cannot show
q147's entry has already been scoped once, in cycle 135, because GSE252378 contains no megakaryoblastic sample and IRX1 expression in AML is an FAB-M7 property independent of the fusion. This beat adds a second scope limit - two independent patients, one institution - and does not revisit the first.
Evidence that argues against this (1)
- The published expression literature does not read IRX1 as an AFF1-partner effect. Symeonidou and Ottersbach define HOXA9/IRX1 as two SUBGROUPS within MLL-AF4 infant ALL, not as a contrast between AFF1 and other partners, and Kuhn's t(4;11)-specific mechanism is IRX1 protein acting ON the MLL-AF4 complex rather than the fusion occupying IRX1. A partner-specificity reading of the occupancy data is this project's, not the field's. Symeonidou V et al. 2021
- E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
IRX1 is expressed exclusively in pro-B cells among the normal stages of B-cell development, and a 2022 survey of TALE-class homeobox genes in BCP-ALL identified aberrant IRX2, IRX3 and MEIS1 - not IRX1 - as the deregulated members, with MEIS1 rather than IRX1 correlating with KMT2A fusions.
Normal and Aberrant TALE-Class Homeobox Gene Activities in Pro-B-Cells and B-Cell Precursor Acute Lymphoblastic Leukemia · Nagel S, Meyer C · International Journal of Molecular Sciences 23:11874 2022What this evidence cannot show
This is an expression study in cell lines and patient cohorts with no chromatin occupancy measurement, so it neither supports nor refutes an occupancy claim directly. Its relevance is that IRX1 is a normal pro-B property, so IRX1 activity in an infant pro-B leukaemia is not by itself evidence of a fusion-driven address.
Evidence that argues against this (1)
- Nagel's survey is of TALE-class genes by expression correlation with subtype markers, and its KMT2A arm is MEIS1. It does not report a systematic IRX1 analysis in KMT2A-rearranged cases, so its silence on IRX1 in that subtype is not a negative result about IRX1 there. Kuhn A et al. 2016
- supports →IRX1 fusion occupancy is confined to KMT2A::AFF1: 0 of 18 samples across five other partners
the headline is a correct measurement of a pattern that is not what chance gives at sample level; it should be restated with its n (two patients, one institution) rather than withdrawn
the class null is the primary test the branch is scored on
the per-gene patient-level p is the condition SURVIVES fails on
the FALLS condition that fired and the measurement that overrides it
scores the hypothesis against a pre-registered gene-scale null; it is not refuted and is not supported at patient level