Applied symmetrically, the criteria that were never applied make IRX1 the weakest member of the class it won - and MEIS1, a canonical fusion target, is indistinguishable from it, so the deposit cannot separate the two
also called q167 apply every criterion to the survivor · IRX1 is 2 patients · MEIS1 is also 2 patients · the criteria cannot discriminate at this n
Over several cycles a six-gene list was cut to one - IRX1 - by removing members one at a time. Each removal used a different test. Running the tests that were never used shows IRX1 would have been the first to go on one of them. But a gene nobody doubts scores exactly the same, so the dataset is simply too small to tell them apart.
The same thing, for a scientist
🔴 **CORRECTION, cycle 154 (see ana-q169-...).** The comparison below - that MEIS1 is indistinguishable from IRX1 on independence - is an ARTEFACT OF THE INSTRUMENT. MEIS1 was scored here with this project gene-body rule, which recovers only 20.96% of the authors calls and gave it 2 samples / 2 patients. **In the authors own caller MEIS1 has 7 samples / 6 patients, the 99.9th percentile**, against IRX1 at 3 / 2 and the 60.5th. The arithmetic below is correct; the conclusion drawn from it is not. **What replaces it.** MEIS1, PROM1, HOXA9, ELF1, GNAQ and PAN3 are all NON-exclusive (5 to 16 non-AFF1 calls), so comparing an exclusive gene against them on exclusivity-adjacent support was the wrong axis. The right comparator is the authors own AFF1-exclusive set: **378 genes, 52.6% of everything called in any AFF1 sample, of which 72.2% rest on ONE patient.** Within that set IRX1 is at the **87.6th percentile**. So IRX1 is a well-supported member of a category that is mostly single-observation noise - weaker than "partner-specific address" and stronger than "indistinguishable from MEIS1". **Both of this entry framings were wrong, in opposite directions.** **The underpowering conclusion survives** - it just rests on the 72.2% single-patient rate rather than on the MEIS1 comparison. Cycle 149 recorded an objection against its own conclusion: **a result reached by serial elimination invites the question of whether the criteria were applied as severely to the survivor.** It was written as an open question and not run. This runs it. **The five criteria that were applied**, each removing a different member: authors'-caller exclusivity (ELF1, GNAQ, PAN3), histone-cluster membership (H3C15), worst margin at 1.01x (H3C15), N/C at background (H3C15), C-occupancy wherever N-occupied (SPAG7). **IRX1 passes all five.** **Four criteria the project uses constantly and never tabulated per member** - recurrence, independent patients, institutions, lineage span - computed symmetrically [E1]: | gene | samples | **patients** | institutions | lineages | |---|---:|---:|---:|---| | **IRX1** | 3 | **2** | 1 | AML/B | | H3C15 | 3 | 3 | 1 | B | | SPAG7 | 3 | 3 | 2 | B | | ELF1 | 5 | 4 | 1 | AML/B | | GNAQ | 5 | 5 | 2 | AML/B | | PAN3 | 7 | 6 | 2 | AML/B | 🔴 **IRX1 is the weakest member of the class it won.** It is **uniquely lowest on independent patients (2, against 3 to 6)**, tied-lowest on recurrence with the two members already removed, and tied-lowest on institutions. **Had independence been the elimination criterion, IRX1 would have gone first.** Its survival is a property of which criteria were chosen, exactly as cycle 149 feared [E1]. 🔵 **But the comparators refuse to let that be the conclusion.** Scored by the identical rule on the same samples [E2]: | canonical fusion target | samples | patients | institutions | |---|---:|---:|---:| | **MEIS1** | 2 | **2** | **1** | | **PROM1** | 2 | **2** | **1** | | HOXA9 | 3 | 3 | 2 | 🟠 **MEIS1 is 2 patients and 1 institution. So is PROM1. So is IRX1.** MEIS1 is a textbook KMT2A-fusion target that nobody doubts. **A criterion that disqualifies IRX1 on independence disqualifies MEIS1 on the same numbers**, which means the criterion is not measuring what it is being asked to measure [E2]. 🟢 **So the verdict is NEITHER, and that is the useful answer.** IRX1's survival is not evidence - it is criterion-dependent. IRX1's weakness is not a refutation - the canonical targets share it. **What this establishes is that GSE252378 is underpowered for the question the last several cycles have been asking**: at 8 index samples collapsing to 6-7 patients across 4 institutions, per-gene independence cannot separate a real target from an artefact. **The elimination sequence was measuring the deposit, not the genes.**
What this does to five cycles of work
🔵 It does not retract the eliminations. The five members were removed on criteria that are about the evidence for each gene - being called in non-AFF1 samples, sitting on the threshold, being C-occupied. Those remain valid reasons and are independent of sample size.
🔴 It retracts the conclusion drawn from the survival. "IRX1 is the only member that survives all four tests" was true and, read as support for IRX1, unjustified. The honest restatement: five members were removed for cause; the sixth was not tested against the criteria on which it is weakest; and on those criteria it is indistinguishable from genes nobody disputes.
🟠 The lineage column is the one place IRX1 differs and it cuts both ways. IRX1 spans AML and B (p318 is the AML sample), as do ELF1, GNAQ and PAN3, while H3C15 and SPAG7 are B-only. Spanning lineages is why IRX1 cannot appear in a lineage-matched class - a liability for that test - but it is also the only member-level evidence of generality across lineage that any of the six has.
🔵 And the comparator choice matters, so it should be criticised. HOXA9, MEIS1 and PROM1 were chosen as canonical targets before their numbers were seen, but they were chosen by me and there are only three of them. A larger comparator set - every gene in the authors' own 1,091-gene table that is called in AFF1 samples - would make the 'indistinguishable' claim quantitative rather than anecdotal, and is the obvious next step.
| script | analyses/q167-apply-every-criterion-to-the-survivor/run.py |
|---|---|
| status | run |
| answers open question from | cycle 149 king log and ana-q163 |
| verdict | NEITHER - survival is criterion-dependent, weakness is shared with canonical targets, and the deposit is underpowered for per-gene independence |
criteria previously applied
[ "authors-caller exclusivity (ELF1, GNAQ, PAN3)", "histone-cluster membership (H3C15)", "margin 1.01x (H3C15)", "N/C at background (H3C15)", "C-occupied wherever N-occupied (SPAG7)" ]
criteria never applied until now
[ "recurrence", "independent patients", "institutions", "lineage span" ]
class — the numbers behind the chart
{
"IRX1": {
"samples": 3,
"patients": 2,
"institutions": 1,
"lineages": [
"AML",
"B"
]
},
"H3C15": {
"samples": 3,
"patients": 3,
"institutions": 1,
"lineages": [
"B"
]
},
"SPAG7": {
"samples": 3,
"patients": 3,
"institutions": 2,
"lineages": [
"B"
]
},
"ELF1": {
"samples": 5,
"patients": 4,
"institutions": 1,
"lineages": [
"AML",
"B"
]
},
"GNAQ": {
"samples": 5,
"patients": 5,
"institutions": 2,
"lineages": [
"AML",
"B"
]
},
"PAN3": {
"samples": 7,
"patients": 6,
"institutions": 2,
"lineages": [
"AML",
"B"
]
}
}canonical comparators — the numbers behind the chart
{
"MEIS1": {
"samples": 2,
"patients": 2,
"institutions": 1
},
"PROM1": {
"samples": 2,
"patients": 2,
"institutions": 1
},
"HOXA9": {
"samples": 3,
"patients": 3,
"institutions": 2
}
}irx1 rank
{
"recurrence": "tied lowest with H3C15 and SPAG7",
"patients": "UNIQUELY lowest (2 against 3-6)",
"institutions": "tied lowest with ELF1 and H3C15"
}measurements
[
{
"quantity": "independent patients supporting IRX1's gene-body occupancy",
"value": 2,
"unit": "patients",
"subject": "uniquely the lowest of the six class members, which span 2 to 6; p279 and p318 are one patient",
"universe": "8 KMT2A::AFF1 samples collapsing to 6-7 patients across 4 institutions"
},
{
"quantity": "independent patients supporting MEIS1, a canonical KMT2A-fusion target, under the identical rule",
"value": 2,
"unit": "patients",
"subject": "MEIS1 2 samples / 2 patients / 1 institution; PROM1 identical; HOXA9 3/3/2",
"universe": "a criterion that disqualifies IRX1 on independence disqualifies MEIS1 on the same numbers"
}
]correction cycle 154
{
"what_was_wrong": "MEIS1 scored with this project rule (2 patients) not the authors caller (6 patients)",
"meis1_authors_caller": {
"samples": 7,
"patients": 6,
"percentile": 99.9,
"non_aff1": 15
},
"irx1_authors_caller": {
"samples": 3,
"patients": 2,
"percentile": 60.5,
"non_aff1": 0
},
"right_comparator": "the authors own 378-gene AFF1-exclusive set, median 1 patient, 72.2% single-patient",
"irx1_percentile_in_exclusive_set": 87.6,
"what_survives": "the underpowering conclusion, now resting on the 72.2% single-patient rate",
"by_entry": "ana-q169-aff1-exclusivity-describes-half-the-table-and-my-meis1-comparison-was-an-artefact-of-the-weaker-instrument"
}- Replace the three hand-chosen comparators with every gene in the authors' 1,091-gene table called in at least one AFF1 sample, and place IRX1's patient count in that distribution. That turns 'indistinguishable from MEIS1' from an anecdote into a percentile.
- State the underpowering as a number. With 8 index samples over 6-7 patients, what effect size could this design detect at all? If the answer is 'only genes occupied in most patients', then the entire exclusive-recurrent framing was outside the deposit's resolution from the start.
- Restate the entries that describe a six-gene class. Five were removed for cause and the sixth is not supported by its survival; the class should be described as what it is.
- Check whether the rule's 20.96% recall against the authors' caller explains MEIS1 and PROM1 scoring 2 of 8. If the rule misses real targets most of the time, comparisons against canonical genes measured by it are weak in both directions.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article
Computed symmetrically across the 8 KMT2A::AFF1 samples for all six members of the gene-body AFF1-exclusive class, the number of N-occupied samples, independent patients (collapsing p279/p318 and R3595/R3652), institutions and lineages spanned are: IRX1 3 samples / 2 patients / 1 institution / AML+B; H3C15 3/3/1/B; SPAG7 3/3/2/B; ELF1 5/4/1/AML+B; GNAQ 5/5/2/AML+B; PAN3 7/6/2/AML+B. IRX1 is uniquely lowest on independent patients, tied-lowest on recurrence with H3C15 and SPAG7, and tied-lowest on institutions with ELF1 and H3C15.
What this evidence cannot show
Patient collapse uses the three replicate pairs recorded in this project's nulls module; if any other two samples are from one patient the counts are too high. Institution assignment is likewise taken from that module and was not re-derived from GEO this session. Only the KMT2A-N flat-floor variant of the occupancy rule was used. These are descriptive counts with no statistical test - 'uniquely lowest' at n=6 with values 2 to 6 is a ranking, not a significant difference.
Evidence that argues against this (2)
- Ranking six genes on four correlated quantities and reporting the one where the survivor comes last is close to guaranteed to find something: with six members and four criteria, some member is worst on each, and IRX1 had a one-in-six chance per criterion of being that member. The finding is only meaningful because the criterion it loses on - independent patients - is the one this project already treats as decisive elsewhere, not because losing on some criterion is surprising.
- Low patient count is partly a consequence of IRX1 being the gene the project selected for, not an independent fact about it. IRX1 was noticed BECAUSE it was called in a small, distinctive set of samples; a gene found that way will systematically have fewer supporting patients than one found by recurrence.
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article
Scored by the identical occupancy rule on the same 8 KMT2A::AFF1 samples, the canonical KMT2A-fusion targets give MEIS1 2 samples / 2 patients / 1 institution, PROM1 2 samples / 2 patients / 1 institution, and HOXA9 3 samples / 3 patients / 2 institutions - so MEIS1 and PROM1 match or fall below IRX1 on every independence criterion on which IRX1 is the weakest class member.
What this evidence cannot show
Three comparators is a small and hand-chosen set. They were selected as canonical KMT2A-fusion targets before their counts were computed, but by me and without pre-registration, and a different three could give a different impression. The argument that the criteria cannot discriminate rests on this comparison and deserves the larger comparator set named in the open questions before it is treated as settled.
Evidence that argues against this (2)
- MEIS1 and PROM1 scoring low here may say more about the rule than about the criteria. Both are well-established fusion targets that this gene-body rule calls in only 2 of 8 samples - a recall problem that this corpus has already documented at 20.96% against the authors' caller. If the rule misses real targets most of the time, then 'IRX1 looks like MEIS1' is a statement about two poorly-measured genes, not evidence that IRX1 is real.
- The symmetric reading is available and less comfortable: if the deposit cannot distinguish IRX1 from MEIS1, it also cannot distinguish IRX1 from a gene that is occupied in two patients by chance. 'Underpowered' protects IRX1 from refutation and denies it support in exactly equal measure, and the entries that treat IRX1 as a partner-specific address should say so.
runs the self-audit that entry named; IRX1s survival is criterion-dependent
IRX1 is the weakest member on independence, the criteria never applied to it
discrimination between partner groups holds; per-gene independence does not
- refutes →IRX1 — Iroquois homeobox 1, the gene that names the HOXA-negative half of infant t(4;11) ALL
2 independent patients, uniquely lowest of the six, and MEIS1 matches it
MEIS1 has 6 patients on the authors caller; the indistinguishability was an instrument artefact
q167 found IRX1 the weakest member of the class it won; the permutation now shows the win itself is within chance