Of the 58 two-patient AFF1-exclusive genes, four - not one - are called exclusive by both instruments, and the surviving IRX1 claim does not beat a patient-level null (p=0.092) [CORRECTED by q171]
also called q170 the two-patient exclusives · IRX1 separates, 1 of 42 · CLEC14A has more calls and no entry · the attention is 111 to 3 to 0 to 0
**CORRECTED (cycle 157, by q171).** The headline below is wrong. At the literal criterion there are FOUR such genes, not one, and 34 of the 42 could never have been one because the second instrument never calls them anywhere. The narrower claim that survives - IRX1 is the only gene called in two or more fusion samples and none of the others - does not beat chance once the null respects that two of IRX1's three samples are the same patient (p=0.092). The occupancy line is finished, as cycle 155 pre-registered it would be on this result. --- Last cycle showed that being 'unique to this fusion' is common - 378 genes have it and most rest on one patient. Narrowing to the 58 with two patients and asking whether a second, independent measurement agrees: only IRX1 passes both. But three genes come close, and the project has written 111 entries about IRX1 and three about all of them combined.
The same thing, for a scientist
> 🔴 **CORRECTED, cycle 157, by `ana-q171-the-two-instrument-intersection-does-not-beat-a-patient-level-null`.** Three defects, all measured: > 1. **The count is 4, not 1.** At the stated criterion (>=1 gene-body AFF1 call and zero gene-body non-AFF1 calls) the hits are **IRX1, ANGPT1, FOXL2 and FOXL2NB**. The table below displays only genes with **two or more** AFF1 calls, so the three single-call genes were never shown - and the prose generalised from the table to all 42. > 2. **The denominator is 8, not 42.** Only 8 of the 42 are called by the gene-body instrument in **any** of the 26 samples; the other **34 fail the criterion by construction** and cannot move the statistic. 'The only one of 42' quoted a denominator that mostly could not fail. > 3. **Three of the four hits are one observation.** ANGPT1, FOXL2 and FOXL2NB have an **identical call vector** - all called in TB051939 alone - and FOXL2/FOXL2NB are **79 bp apart** on chr3. > > **What survives:** IRX1 is genuinely the only gene with >=2 AFF1 calls and zero non-AFF1 calls. **What it is worth:** at sample level p=0.031, but IRX1's three calls are p279, p318 and p454 and **p279/p318 are the same patient** - which is why IRX1 is in a set defined as *two*-patient. Under a patient-level null (7 of 24 patients permuted) **every threshold gives p = 0.092-0.112**. Nothing reaches 0.05. Cycle 155 wrote the acceptance condition in advance: *"if the permutation shows one hit is the expected number, the line is finished - and I should say so rather than find a sixth criterion."* --- Cycle 154 named this as **the one remaining route to a positive result** on the occupancy line: *"cluster the 58 two-patient exclusive genes against IRX1. If IRX1 is unremarkable among them the line is finished; if it separates, that is the first positive discrimination in several cycles."* 🔵 **The patient pairs are spread, so the set is not one sample's idiosyncrasy.** Eight distinct pairs across the 58; the largest, `p279+p454`, carries 15 (25.9%) and is IRX1's. So IRX1's support is the most common pairing rather than a peculiar one - which neither helps nor hurts it [E1]. 🟢 **On cross-instrument agreement, IRX1 separates.** The gene-body rule is an independent measurement from the authors' peak caller - it recovers only **20.96%** of their calls, so agreement is informative. Of the 58, **42 are in the gene-body table** and only **7 (16.7%)** are called by it in any AFF1 sample. Four have two or more, and the exclusivity check separates them [E2]: | gene | authors | gene-body AFF1 | gene-body **non-AFF1** | both instruments exclusive? | |---|---|---:|---:|---| | **IRX1** | p279+p318+p454 | 3 | **0** | **YES** | | CLEC14A | p279+p454 | **4** | 2 | no | | ROMO1 | p247+p318 | 3 | 1 | no | | DUSP6 | p318+p454 | 2 | 2 | no | **IRX1 is the only one of the 58 that both instruments call exclusive.** CLEC14A has *more* AFF1 calls, but the gene-body rule also finds it in `R3595` and `p186`; ROMO1 and DUSP6 likewise fail. **That is a genuine separation and the first positive discrimination this line has produced in several cycles** [E2]. 🔴 **And the same table is an indictment of how the attention was allocated.** Named in node files: **IRX1 111, CLEC14A 3, ROMO1 0, DUSP6 0.** `is-symbol-covered.py` returns **no match** for all three. **CLEC14A has strictly more AFF1 gene-body calls than IRX1 and 3 mentions against 111** [E3]. Three genes with two-instrument support at or near IRX1's grade have essentially no coverage, and the PI's standing instruction is to map every interactor, one entry per thing. 🟠 **How much the separation is worth, stated honestly.** IRX1 was *already known* to be gene-body-exclusive and recurrent - that is the q154 class it belongs to - so finding it again here is not a surprise. **The non-circular content is the rate: 1 of 42.** The double criterion is stringent, and no other two-patient authors-exclusive gene meets it. But **1 of 42 is a singleton**, and a singleton intersection between two sets is what chance often produces.
What q171 measured, and why the sample-level p-value was the wrong one
The permutation this entry asked for was run in cycle 157 with the authors' 42 held fixed and the gene-body instrument's AFF1 labels permuted, 20,000 draws, seed 20260924.
| null | unit | t | observed | null mean | P(>=obs) |
|---|---|---|---|---|---|
| sample-level | gene | 1 | 4 | 1.259 | 0.090 |
| sample-level | gene | 2 | 1 (IRX1) | 0.032 | 0.031 |
| patient-level | gene | 1 | 4 | 1.269 | 0.092 |
| patient-level | gene | 2 | 1 (IRX1) | 0.099 | 0.092 |
| patient-level | distinct call vector | 1 | 2 | 0.684 | 0.112 |
| patient-level | distinct call vector | 2 | 1 (IRX1) | 0.099 | 0.092 |
The only thing separating 0.031 from 0.092 is patient structure. A sample-level null may hand a random gene three calls from three different patients, which IRX1 does not have: its three calls are two patients. The sample-level figure is anticonservative and should not be quoted.
Note the threshold behaviour. The verdict moves with the cutoff - 0.090 at t>=1 and 0.031 at t>=2 under the sample-level null - which is this project's standing warning that a result tracking its threshold is a result about the threshold. Under the patient-level null the curve is flat at ~0.09 and the warning no longer applies, because there is nothing to select: no threshold reaches significance.
Scripts: analyses/q171-permute-the-two-instrument-intersection/{run,detail,thresh,patient}.py.
Why this is not the replication it resembles
🔴 Both instruments read the same 27 libraries. The authors' caller and this project's gene-body rule are different methods applied to one experiment - same patients, same antibodies, same sequencing. A sample-level artefact appears in both. This corpus already recorded that point for the six-gene class and it applies unchanged here: cross-method agreement within one experiment is weaker than it sounds.
🟠 Sixteen of the 58 are not in the gene-body table at all and are therefore unscored, not negative. If any of them would have passed, the 1-of-42 becomes 1-of-58 or worse, and the claim that IRX1 is the only one is a claim about the 42.
🔵 The cheapest thing that would make this stick is the one thing this deposit cannot provide: a second cohort. Three genes now have IRX1-comparable evidence within GSE252378, and distinguishing four candidates on 8 samples is exactly the underpowering cycle 152 established. What CLEC14A, ROMO1 and DUSP6 earn is entries, not a ranking.
| script | analyses/q170-is-irx1-unremarkable-among-the-two-patient-exclusives/run.py |
|---|---|
| status | run |
| answers open question from | cycle 154 king log and ana-q169 |
| n two patient exclusive | 58 |
| irx1 pair | p279+p454 |
| irx1 pair share PCT | 25.9 |
| n in gene body table | 42 |
| n with any gene body call | 7 |
| PCT with any | 16.7 |
| verdict | IRX1 separates - the only one of 42 scored that both instruments call exclusive - but 1 of 42 is a singleton and no permutation was run on it |
patient pairs — the numbers behind the chart
{
"p279+p454": 15,
"p179+p247": 11,
"TB986292+p454": 9,
"p247+p279": 8,
"TB051939+p247": 5,
"p247+p454": 4,
"p179+p279": 3,
"p179+p454": 3
}top four — the numbers behind the chart
{
"IRX1": {
"authors": [
"p279",
"p318",
"p454"
],
"gb_aff1": 3,
"gb_nonaff1": 0,
"both_exclusive": true
},
"CLEC14A": {
"authors": [
"p279",
"p454"
],
"gb_aff1": 4,
"gb_nonaff1": 2,
"gb_nonaff1_samples": [
"R3595",
"p186"
],
"both_exclusive": false
},
"ROMO1": {
"authors": [
"p247",
"p318"
],
"gb_aff1": 3,
"gb_nonaff1": 1,
"gb_nonaff1_samples": [
"R3595"
],
"both_exclusive": false
},
"DUSP6": {
"authors": [
"p318",
"p454"
],
"gb_aff1": 2,
"gb_nonaff1": 2,
"gb_nonaff1_samples": [
"A80956",
"A93692"
],
"both_exclusive": false
}
}corpus coverage — the numbers behind the chart
{
"IRX1": 111,
"CLEC14A": 3,
"ROMO1": 0,
"DUSP6": 0
}is symbol covered
{
"CLEC14A": "no match",
"ROMO1": "no match",
"DUSP6": "no match"
}measurements
[
{
"quantity": "two-patient AFF1-exclusive genes that both instruments call exclusive",
"value": 1,
"unit": "of 42 scored (IRX1)",
"subject": "the authors' peak caller and this project's gene-body rule, which recovers 20.96% of their calls",
"universe": "58 two-patient exclusives, 42 in the gene-body table, 7 with any gene-body AFF1 call, 4 with two or more"
},
{
"quantity": "node files naming CLEC14A, which has more AFF1 gene-body calls than IRX1",
"value": 3,
"unit": "node files against IRX1's 111",
"subject": "ROMO1 and DUSP6 are named in zero; is-symbol-covered returns no match for all three",
"universe": "CLEC14A has 4 KMT2A::AFF1 gene-body calls against IRX1's 3, but 2 non-AFF1 calls against IRX1's 0"
}
]- RESOLVED, cycle 157 (q171): the permutation was run. Patient-level p = 0.092 at every threshold; the observed count at the literal criterion is 4, not 1, and the effective denominator is 8, not 42. The occupancy line is finished on its own pre-registered acceptance condition. The remaining open questions below are about the OTHER genes and stand on their own.
- ANGPT1, FOXL2 and FOXL2NB are exclusive by the literal criterion and have no entry either. FOXL2/FOXL2NB are 79 bp apart and ANGPT1 shares their exact call vector - all three rest on the single sample TB051939, so they are one observation and a weak one. Record that judgement rather than leaving them unmentioned, which is the error this entry already made about CLEC14A, ROMO1 and DUSP6.
- Write CLEC14A, ROMO1 and DUSP6 as entries, or record explicitly why not. They have two-instrument occupancy support at or near IRX1's grade and between them 3 node-file mentions; the PI's instruction is one entry per thing, and a reasoned refusal is also an acceptable outcome provided it is written down.
- Score the 16 of 58 that are absent from the gene-body interval table. They are unscored, not negative, and they are the difference between 'the only one of 42' and 'the only one of 58'.
- Check whether p279 and p454 are simply the deepest AFF1 libraries. Their pair carries 26% of the two-patient set, and library depth was not controlled for.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article
The 58 AFF1-exclusive genes in Janssens' table that rest on exactly two independent patients are supported by 8 distinct patient pairs, the commonest being p279+p454 with 15 genes (25.9%), which is also IRX1's pair; the remaining pairs carry 11, 9, 8, 5, 4, 3 and 3 genes, so the set is not the idiosyncrasy of a single sample pairing.
What this evidence cannot show
Patient collapse uses this project's three recorded replicate pairs; IRX1's three samples p279, p318 and p454 collapse to two patients because p279 and p318 are one child. With 8 AFF1 samples collapsing to 7 patients there are only 21 possible pairs, so 8 observed pairs across 58 genes is not a strong statement about spread.
Evidence that argues against this (1)
- The commonest pair carrying 26% of the set is compatible with a sample-quality effect rather than biology: if p279 and p454 are the two deepest or cleanest libraries, genes called in exactly two samples will preferentially be called in those two. Library depth was not controlled for here.
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.journal-article
Of the 58 two-patient AFF1-exclusive genes, 42 appear in this project's gene-body interval table and only 7 (16.7%) are called by the gene-body rule in any KMT2A::AFF1 sample; four have two or more such calls - CLEC14A 4, IRX1 3, ROMO1 3, DUSP6 2 - and of those only IRX1 has zero gene-body calls among the 18 non-AFF1 samples, CLEC14A being called in R3595 and p186, ROMO1 in R3595 and DUSP6 in A80956 and A93692. IRX1 is therefore the only member of the 58 that both the authors' caller and this project's independent gene-body rule call AFF1-exclusive.
What this evidence cannot show
IRX1 was already known to be gene-body-exclusive and recurrent - it is a member of the q154 six-gene class - so recovering that here is not independent information; the non-circular content is the RATE, 1 of 42. Sixteen of the 58 are absent from the gene-body interval table and are unscored rather than negative, so 'the only one' is a statement about the 42. Only the KMT2A-N flat-floor variant of the rule was used.
Evidence that argues against this (2)
- Both instruments read the same 27 libraries of one deposit, so this is cross-method agreement within a single experiment, not replication. A sample-level artefact - a mislabelled library, a batch effect, a copy-number gain at 5p15 - would produce agreement between the two methods just as readily as biology would.
- A singleton intersection between two sets is a weak result. With 42 genes scored and a gene-body exclusivity rate of roughly 5% among them, one hit is close to the expectation, and no permutation was run to ask how often the intersection is empty or singleton by chance. That test is one loop and was not done.
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Of the four genes with two or more gene-body calls among the 58, coverage in this graph is IRX1 111 node files, CLEC14A 3, ROMO1 0 and DUSP6 0, and scripts/is-symbol-covered.py returns 'no match' for CLEC14A, ROMO1 and DUSP6 on a plain invocation - despite CLEC14A having more KMT2A::AFF1 gene-body calls (4) than IRX1 (3).
What this evidence cannot show
A no-match from is-symbol-covered.py is explicitly NOT proof of a gap - the tool's own footer says so, and this project has recorded that 225 entries across five prefixes are unreachable by its id arm. The three genes may be discussed inside a complex or mechanism entry that names neither the symbol nor a decline phrase. The 111 figure for IRX1 counts every mention including passing ones.
Evidence that argues against this (2)
- The attention asymmetry is not by itself evidence of error. IRX1 was pursued because of a published expression subgroup and a prior mechanistic literature, not because of its occupancy counts, so 111 entries reflect a different and legitimate reason for interest. CLEC14A having more gene-body calls does not make it a better candidate for a beat - it is an endothelial C-type lectin with no leukaemia prior at all.
- Writing entries for CLEC14A, ROMO1 and DUSP6 on this evidence risks exactly the error q152 declined to make when it refused to write twelve gene entries on a statistic shown to be at chance. Their support is two-patient occupancy in one deposit, which cycle 154 established is the majority category.
runs the clustering that entry named as the last route to a positive
- supports →IRX1 — Iroquois homeobox 1, the gene that names the HOXA-negative half of infant t(4;11) ALL
the only one of 42 that both instruments call exclusive; 1 of 42 is still a singleton
same two instruments, same caution: cross-method within one experiment is not replication
the permutation q170 asked for: its count of 1 is 4, its denominator of 42 is 8, and p=0.092
records the refusal q170 left owed on ANGPT1/FOXL2/FOXL2NB, on corrected grounds