IRX1's own three-sample pattern is p = 0.022 naively and 0.077-0.176 once patients, institution or lineage are held fixed - the class it belongs to is significant, IRX1 alone is not
also called q154 T3 · the per-gene test for IRX1
IRX1 has the fusion protein sitting on it in three of the eight AFF1 leukaemias and none of the other eighteen, which taken on its own looks unlikely by chance - about one in fifty. But two of those three samples are the same child sampled twice and all three come from one hospital, and correcting for that drops the odds to between one in six and one in thirteen.
The same thing, for a scientist
Under the deposit paper's own per-sample target calls, IRX1 is called in `p279`, `p318`, `p454` and in no other sample of the 26 scored. A label permutation on that single row gives **p = 0.0219** unstratified (3 of 3 calls inside an 8-of-26 index arm). Three declared corrections each cost roughly a factor of four to eight: **exact institution-stratified permutation over all 4,410 assignments p = 0.1190**; **B-lineage-only (16 samples, 7 AFF1) p = 0.1756**; **patient-level (24 patients, 7 AFF1, 2 of them IRX1-positive) p = 0.0772** [E1]. None clears 0.05. **The honest reading is that IRX1 individually is a single observation of 2 independent patients, and the thing that is statistically established is the class** - 64 AFF1-exclusive recurrent target genes against a stratified null of 8.05, p_upper 0.0027 - **of which IRX1 is one member** [E2]. A p of 0.02 on one gene chosen out of a 1,091-gene table is also, on its own, exactly what 22 genes would show by chance.
The four tests and what each one removes
| test | index arm | IRX1 calls inside / outside | p | what it removes |
|---|---|---|---|---|
| sample level, unstratified | 8 of 26 | 3 / 0 | 0.0219 | nothing |
| patient level | 7 of 24 | 2 / 0 | 0.0772 | p279/p318 are one patient |
| institution-stratified, EXACT | 5 of 9 Texas + 3 of 7 St Jude | 3 / 0 | 0.1190 | all three IRX1-positive samples are Texas Children's p-series |
| B-lineage only | 7 of 16 | 2 / 0 | 0.1756 | the eight Fred Hutch comparators are AML and carry no AFF1 |
The institution test is exhaustive, not sampled: the whole permutation space under that constraint is 4,410 assignments and every one was evaluated, so the 0.1190 is exact and carries no Monte-Carlo error 1.
This is not a refutation of the IRX1 call
A null says what a pattern is worth as evidence for partner-specificity. It says nothing about whether the fusion is on IRX1 in p279, p318 and p454. Two independent calling pipelines agree that it is, the signal is 6.08x and 6.35x an average promoter in two of them, and the deposit paper states the conclusion in its own discussion. What fails is the inference from three samples to "the IRX1 address requires AFF1", because three samples from one institution that are two patients cannot carry that inference on their own - and, separately, because the class-level result shows the same deposit contains dozens of equally partner-exclusive genes nobody has named 2.
Why the direction of each correction was declared first
Each of the three corrections could only move the p-value up, and that was written down before any was run 3: collapsing a pair removes an observation from the numerator; stratifying by institution removes draws in which the index arm could not have been the p-series; restricting to B-lineage removes comparators that were never candidates. A correction whose direction is known in advance is not a fishing expedition, and a correction whose direction is not written down is one. The escape condition that fired here - the verdict differing between sample level and patient level - was pre-declared as U5.
| script | analyses/q154-gene-scale-occupancy-null/authors_null2.py |
|---|---|
| status | run |
| result summary | IRX1 alone: p 0.0219 unstratified, 0.0772 patient level, 0.1190 exact institution-stratified, 0.1756 B-lineage only. None clears 0.05. The 64-member class it belongs to does. |
inputs
[ "Janssens 2024 Supplementary Data 2 (41467_2024_53399_MOESM4_ESM.xlsx)", "analyses/q144-unbounded-magnitude/manifest.json (institution and lineage labels)" ]
measurements — the numbers behind the chart
[
{
"quantity": "label_permutation_p_for_one_gene",
"subject": "IRX1 called in 3 of 8 KMT2A::AFF1 samples and 0 of 18 non-AFF1, unstratified",
"value": 0.0219,
"units": "one-sided permutation p, +1 corrected",
"n": 100000,
"universe": "26 GSE252378 KMT2A-fusion patient samples, authors' own oncoprotein-target calls",
"null": 0.5,
"note": "this is the value that does NOT survive any of the three declared corrections"
},
{
"quantity": "label_permutation_p_for_one_gene_institution_exact",
"subject": "IRX1's same pattern under an exhaustive institution-stratified permutation",
"value": 0.119,
"units": "exact permutation p",
"n": 4410,
"universe": "all 4,410 assignments keeping 5 of 9 Texas Children's and 3 of 7 St Jude samples in the index arm",
"null": 0.5,
"note": "exhaustive; no Monte-Carlo error. All three IRX1-positive samples are Texas Children's."
},
{
"quantity": "label_permutation_p_for_one_gene_patient_level",
"subject": "IRX1 called 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": 100000,
"universe": "24 independent patients after collapsing p279/p318, 148752/152985 and R3595/R3652",
"null": 0.5,
"note": "3 of 8 samples is 2 of 7 patients"
},
{
"quantity": "label_permutation_p_for_one_gene_b_lineage_only",
"subject": "IRX1 restricted to B-lineage samples: 2 of 7 AFF1, 0 of 9 non-AFF1",
"value": 0.1756,
"units": "one-sided permutation p, +1 corrected",
"n": 100000,
"universe": "the 16 B-lineage GSE252378 samples; p318 is classified AML and drops out",
"null": 0.5,
"note": "the eight Fred Hutch comparators are paediatric AML and carry no AFF1 sample"
}
]- Any biomarker or stratification built on IRX1 occupancy currently rests on two patients from one hospital. A single institution's sample handling - and these are the five highest-signal libraries in the deposit - is a sufficient alternative explanation for the whole pattern until a second centre reproduces it.
- An eighth independent KMT2A::AFF1 patient from a different institution, IRX1-positive, would move the patient-level p from 0.077 to roughly 0.02 and the institution-stratified p out of its current range. Which cohorts hold an untreated KMT2A::AFF1 blast with deposited oncoprotein CUT&RUN outside Texas Children's?
- Is p318 B-lineage or AML for the purpose of a lineage-matched null? It is the AML arm of a lineage switch, so both classifications are defensible and they give p = 0.176 and a lower value respectively. A pre-declared rule for lineage-switching pairs is missing from this project.
- Of the 64 class members, how many have an unstratified p at or below IRX1's 0.0219? If most do, IRX1's individual p adds nothing at all beyond membership.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Label permutation on IRX1's own row of the published GSE252378 target table gives p = 0.0219 unstratified over 26 samples, 0.0772 at patient level over 24 patients, 0.1190 under an exhaustive institution-stratified permutation over all 4,410 assignments, and 0.1756 restricted to the 16 B-lineage samples.
Q154 label permutation over the published GSE252378 oncoprotein-target call table · Iqbal Lab oncofusion knowledge graph, beat Q154 · this project 2026What this evidence cannot show
The institution labels come from the GEO sample descriptions in the q144 manifest (Texas Children's, St Jude, Fred Hutch, Seattle Children's) and were read after the pre-registration was written. The B-lineage restriction classifies p318 - the AML arm of the p279 lineage switch - as AML, which removes one of IRX1's three calls; classifying it as B-lineage instead would give 3 of 8 in a 17-sample restriction and a lower p. That choice is stated because it changes the number.
Evidence that argues against this (1)
- All four tests condition on IRX1 having been selected for testing, and IRX1 was selected because it is already this project's headline. A gene picked out of a 1,091-row table on the strength of an unstratified p of 0.02 is expected to appear about 22 times by chance in that table. Janssens DH et al. 2024
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The class IRX1 belongs to - genes called in >=3 of the 8 AFF1 samples and 0 of the 18 non-AFF1 - numbers 64 against an exact institution-stratified permutation null of 8.05, p_upper 0.0027, so the class carries the statistical weight that the single gene does not.
Q154 label permutation over the published GSE252378 oncoprotein-target call table · Iqbal Lab oncofusion knowledge graph, beat Q154 · this project 2026What this evidence cannot show
A significant class does not make every member significant, and the class is 30% HIST1 cluster genes. Membership is evidence that IRX1 is not alone, not evidence that IRX1 in particular is a partner-specific address.
Evidence that argues against this (1)
- The class result is computed on the authors' own nominated gene set. If their peak caller systematically nominates more intervals in high-signal libraries - and the five highest-call-rate libraries are five of the eight AFF1 samples - then the class itself inherits that bias. The call-rate-matched permutation addresses the per-sample rate but not the composition of the 1,091-gene universe. NCBI Gene Expression Omnibus 2024
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The pre-registration required that the verdict be stable across floors and across sample versus patient level, named U5 as the escape condition when it is not, and required the direction of every control to be written down in advance.
Q154 pre-registration, written before computation · Iqbal Lab oncofusion knowledge graph, beat Q154 · this project 2026What this evidence cannot show
Written and executed by the same agent in one session.
Evidence that argues against this (1)
- The pre-registration's own control table names institution only inside a general list; the exact 4,410-assignment enumeration and the B-lineage restriction were specified after the per-sample institution and lineage strings were read from the deposit metadata. A reader should treat those two stratifications as post-hoc, which is the direction that weakens rather than strengthens the corrections reported here. Iqbal Lab oncofusion knowledge graph et al. 2026
IRX1 p 0.0219 to 0.1756 reproduced; none clears 0.05
independent route to the same conclusion about IRX1 individually
- refutes →IRX1 fusion occupancy is confined to KMT2A::AFF1: 0 of 18 samples across five other partners
the 0-of-18 is a real measurement but IRX1's own pattern carries p = 0.077 at patient level and 0.119 institution-stratified, so it cannot on its own establish that the address requires AFF1
the class-level result is what IRX1's membership borrows its weight from
scores the hypothesis's central pattern against four label-permutation nulls; it survives none at 0.05 on its own
- related to →IRX1 — Iroquois homeobox 1, the gene that names the HOXA-negative half of infant t(4;11) ALL
IRX1's partner-exclusive occupancy pattern, scored against its own null at sample, patient, institution and lineage level
the per-gene patient-level p is the condition SURVIVES fails on