Oncofusion

start / graph / ana-q153-q151s-irx2-negative-survives-attribution-because-a-null-over-a-superset-entails-a-null-over-its-subset

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q153-can-an-irx2-promoter-measurement-be-attributed

Attribution is asymmetric: q151's IRX2 negative survives a window that is 89.2% someone else's gene, and a positive in the same window would not have

also called IRX2 negative attribution · asymmetry of attribution · q151 verdict

Finding nothing in a stretch of DNA that contains IRX2's on-switch really does mean nothing was at IRX2's on-switch, no matter which other genes share that stretch. Finding something there would not have told you which gene it belonged to. So the original negative result stands, and the near-miss that came with it does not.

The same thing, for a scientist

The pre-registered prediction for this beat was that attribution is **asymmetric**, and the geometry bears it out. The `IRX2` row of this project's promoter table measures `chr5:2,750,677-2,752,677`, which **is** IRX2's own strand-correct 2 kb promoter window — verified against two independently fetched annotations, including the minus-strand check that its TSS is the **higher** coordinate of the gene body [E1]. Whatever else occupies that interval, **IRX2's promoter is a subset of the region measured**, so *no signal anywhere in the window* entails *no signal at IRX2's promoter*. **`ana-q151-irx2-is-not-fusion-occupied-at-promoter-scale-in-gse252378` — 0 of 27 samples, against a null of 2.73 for an arbitrary promoter — therefore stands unchanged as a statement about IRX2** [E1][E2]. The converse fails: a positive would have been a statement about 2 kb of chromosome 5 with two equally good claimants, and the general measurement of how often such a positive can be told apart from its neighbour's is **8.7%** [E3]. **What does inherit the caveat is the near-positive.** q151 named `p454` honestly at 3.81 in KMT2A-N1 and 1.898 in N2, missing the gate by 0.102 share units. Had that sample cleared the gate, the resulting positive would have been **at most 80.3% attributable and possibly 0% attributable** to IRX2 [E4]. **My own pre-registered falsifier did not fire, and saying why matters.** I pre-declared that if the negative rested on a ratio to a floor rather than on absolute absence, a shared window could dilute a real IRX2 peak below threshold. It cannot: the window is `TSS ± 1000` regardless of what the neighbour is, so **the neighbour's existence changes the interpretation of the measurement and not the measurement** [E1]. The dilution risk is real but it is a property of using a 2 kb mean at all, and it applies identically to every clean promoter in the same table.

Go deeper
The argument, set out so it can be attacked

Let W = chr5:2,750,677-2,752,677, the interval measured. Let P_IRX2 ⊂ W be whatever the true IRX2 promoter is — it is inside W by construction, because W is centred on IRX2's TSS.

  1. The measurement reports mean coverage over W and reports it below threshold in 27 of 27 samples.
  2. Coverage is non-negative, so a low mean over W bounds the total signal over W, and therefore over P_IRX2.
  3. Hence: not occupied over W ⟹ not occupied over P_IRX2. The presence of IRX2-DT's promoter inside W is irrelevant to this step; it would be irrelevant if there were ten genes in there.
  4. The reverse inference — occupied over W ⟹ occupied over P_IRX2 — is invalid, because the signal may sit in W \ P_IRX2, which is most of W.

Where the argument can be attacked, and both should be checked by a reader:

  • Step 1 depends on the window being in the right place. A minus-strand gene's TSS is the higher coordinate. If the table had used chrstart, the window would sit at ~2,735,634, about 16 kb away, and the negative would be about empty sequence. It does not: the table stores 2751677 = refGene txEnd, confirmed against NCBI esummary's chrstart 2751676 > chrstop 2735634.
  • Step 2 is about a MEAN, not a maximum. A very narrow, very tall peak confined to P_IRX2 could in principle be averaged down over 2,000 bp. That is a genuine false-negative risk — but it is a property of the 2 kb convention and not of the overlap, and it applies equally to HOXA9 and MEIS1, which the same rule calls occupied in 21 and 24 of 27. The positive controls firing is what makes the dilution objection weak here.
The pre-registered branch that fired: NEITHER, trigger 3

The pre-registration offered ATTRIBUTABLE, NOT ATTRIBUTABLE, and a NEITHER branch with three declared triggers. Trigger 3 fired: the answer is scale-dependent.

scaleseparable?does this deposit have it?
2 kb promoter meanno — 89.2% shared, calls agree 91.3% of the time at such pairsyes, and it is all q151 used
gene bodyyes — the two bodies are disjoint, 5,833 bp and 6,686 bp with a 215 bp gapno such measurement exists
base-resolution coverageyes in principle — CUT&RUN is a high-resolution assayin the bigWigs, never extracted
strandno — CUT&RUN is a DNA-fragment readout with no transcriptional strandnot available at all, ever

So neither top-line branch is correct on its own. The promoter-scale measurement is not attributable; the scales that would be attributable were not measured. Declaring this possibility in advance is the reason it can be reported as a result rather than as a rescue.

What to change in the corpus, and what not to

Do not relabel q151's headline. "IRX2 is not fusion-occupied at promoter scale in GSE252378" is correct as written and does not need to become "the IRX2/IRX2-DT locus is not occupied" — although that second sentence is also true, and is a free extra finding: the same 27 zeros are a negative for IRX2-DT as well, which is the one place a shared window pays a dividend.

Do add the caveat to the near-positive. Any future sentence of the form "p454 comes close at IRX2" must carry that at least 80.3% of that signal is shared with IRX2-DT.

Do not extend the argument to any positive. If a later beat scores IRX2 under a laxer rule, or at a different window, and gets a positive, that positive is not an IRX2 result without base-resolution support.

Facts and figures
statusrun
prereganalyses/q153-can-an-irx2-promoter-measurement-be-attributed/PREREGISTRATION.md
branch firedNEITHER / UNINFORMATIVE, pre-declared trigger 3: the answer is scale-dependent - promoter scale NOT attributable, gene-body scale attributable but never measured
verdict on q151STANDS UNCHANGED as an IRX2 statement. The negative is over a superset of IRX2's promoter and therefore entails the negative over IRX2's promoter.
prereg falsifier outcomeDID NOT FIRE. I pre-declared that a floor-referenced ratio rule could let a shared window dilute a real IRX2 peak below threshold. The window is TSS+/-1000 irrespective of the neighbour, so the neighbour cannot dilute anything; the dilution risk is a property of the 2 kb convention and applies equally to every clean promoter in the same table.
separability by scale
{
  "2kb_promoter_mean": "NOT separable (89.2% shared; calls agree 91.3% at such pairs)",
  "gene_body": "separable in principle (bodies disjoint, 215 bp apart) - NO SUCH MEASUREMENT EXISTS in this project",
  "base_resolution": "separable in principle - in the bigWigs, never extracted; no bigWig was downloaded in this beat",
  "strand": "NOT separable, ever - CUT&RUN is a DNA-fragment readout with no transcriptional strand"
}
measurements
[
  {
    "quantity": "n_samples_occupied",
    "subject": "IRX2 2 kb promoter, GSE252378, re-stated from q151 and unchanged by attribution",
    "value": 0,
    "units": "samples",
    "n": 27,
    "universe": "27 KMT2A-fusion samples; window chr5:2,750,677-2,752,677, which contains IRX2's promoter as a subset",
    "null": 2.73,
    "note": "a null over a superset entails a null over its subset, so the zero is attributable to IRX2 even though the window is 89.2% IRX2-DT"
  },
  {
    "quantity": "max_shared_fraction_of_a_hypothetical_positive",
    "subject": "p454 KMT2A-N1 IRX2 window signal, upper bound on the part shared with IRX2-DT",
    "value": 80.3,
    "units": "percent",
    "n": 1,
    "universe": "p454 KMT2A-N1; complement of the 19.7% one-sided lower bound from two-window deconvolution",
    "null": 98.2,
    "note": "null is the complement of the 0.018 median attributable fraction at shared pairs; p454 is a better-than-typical case and still mostly shared"
  }
]
How this could bite us back
  • The asymmetry is a licence to trust negatives and distrust positives at shared promoters, and that is a dangerous shape for a target-nomination pipeline: it means the errors all run in the direction of nominating a gene, never in the direction of clearing one. A screen built on this data will produce false targets and no false clears.
  • Equally, the free second negative cuts both ways: declaring IRX2-DT unoccupied on the strength of IRX2's zeros is correct, but it would be wrong to then treat IRX2-DT as 'checked' at any other scale, since nothing else about it has been measured at all.
Still open
  1. Is IRX2 occupied at GENE-BODY scale? This is the scale that WOULD be attributable and no such measurement exists anywhere in this project. The interval is chr5:2,745,844-2,751,677 and the machinery is regions147.py. It is the cheapest remaining question at this locus.
  2. Does the subset argument survive a narrow-peak objection? A 200 bp summit at IRX2's TSS could be averaged down over 2 kb. The positive controls firing at 21 and 24 of 27 make it unlikely but do not rule it out, and only base-resolution coverage would.
  3. How many other bounded negatives in this corpus are at shared windows, and therefore get a second gene's negative for free? The census names 4,191 shared windows; nobody has crossed that list against the corpus's absence entries.
Every source behind this entry (4)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    The IRX2 row of this project's promoter table measures chr5:2,750,677-2,752,677, which is IRX2's strand-correct 2 kb TSS window (minus strand, TSS at the higher coordinate 2,751,677, confirmed by refGene txEnd and by NCBI esummary chrstart 2751676 > chrstop 2735634), so IRX2's promoter is a subset of the interval measured and a below-threshold mean over that interval bounds the signal over IRX2's promoter.

    NCBI Gene esummary for GeneID 153572 (IRX2) and UCSC hg38 refGene, cross-checked against C:\Temp\q145irx1\promoters.tsv · NCBI Gene / UCSC Genome Browser / Iqbal Lab oncofusion knowledge graph beat Q153 · NCBI Gene 2026
    What this evidence cannot show

    The subset argument is logical, not empirical: it holds for any non-negative coverage measure over an interval containing the feature of interest. It does NOT hold if the true IRX2 promoter extends outside TSS+/-1000, which is possible - a distal element 3 kb upstream would be outside the window and outside the claim. It also does not hold against a narrow-peak dilution objection, which is addressed separately by the positive controls firing at 21 and 24 of 27 under the same rule.

    Evidence that argues against this (1)
    • A 2 kb MEAN can in principle average away a narrow peak confined to the true promoter, so 'not occupied over W' bounds the total but not the maximum. If the real question is 'is there a summit at IRX2's TSS', the negative is weaker than the subset argument makes it sound, and only base-resolution coverage answers it. Skene PJ et al. 2017
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    q151 reports IRX2's strand-correct 2 kb promoter as occupied in 0 of 27 GSE252378 samples under the q147 two-antibody rule, with HOXA9 at 21 of 27 and MEIS1 at 24 of 27 as positive controls in the same libraries, and an expectation of 2.73 of 27 for an arbitrary promoter under the same permissive rule.

    IRX2 is not fusion-occupied at promoter scale in GSE252378: 0 of 27 samples, against a null of 2.73 · Iqbal Lab oncofusion knowledge graph, beat Q151 · this project 2026
    What this evidence cannot show

    This is a re-statement of another beat's result, read from its entry and its results file, not independently re-run in full. What THIS beat independently re-derived is that the window is in the right place and that the IRX2 and C5orf38 rows are identical to four decimal places in 18 of 27 samples with neither ever occupied. The 27 samples come from 24 patients. No spike-in exists in the deposit.

    Evidence that argues against this (1)
    • The same zeros are simultaneously a negative for IRX2-DT, which q151's own gene-irx2-dt entry records as 0 of 27. That is not a weakness: it is the one benefit of a shared window, and it means the corpus gets two bounded negatives for the price of one. Iqbal Lab oncofusion knowledge graph et al. 2026
  3. E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    A positive call at a window sharing at least half its sequence with a neighbour's is accompanied by the identical call at the neighbour 91.3% of the time in this deposit, against 4.3% for random promoter pairs under the same rule, so a hypothetical IRX2 positive would not have named its gene.

    q153 promoter-call discriminability over overlapping window pairs, GSE252378 · Iqbal Lab oncofusion knowledge graph, beat Q153 · this project 2026
    What this evidence cannot show

    The random-pair null is unmatched on baseline occupancy, which inflates the contrast; the measured 8.7% side is unaffected. This is a statement about the aggregate behaviour of the rule at shared windows, applied here to a hypothetical IRX2 positive that does not exist - it is an argument about what a positive WOULD have meant, not a measurement of one.

    Evidence that argues against this (1)
    • Because no IRX2 positive exists in these 27 samples, the whole positive branch of this argument is counterfactual. A reader entitled to be sceptical can say the attribution problem at IRX2 has cost this corpus nothing so far, and they would be right - the cost is prospective and lands on whoever scores this locus next. Iqbal Lab oncofusion knowledge graph et al. 2026
  4. E4moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    In p454, the one sample q151 named as a near-positive, at most 80.3% of the IRX2 window's KMT2A-N1 signal could be shared with IRX2-DT and at least 19.7% comes from sequence IRX2-DT's window does not contain, so a positive at this sample would have been at best one-fifth attributable to IRX2 by arithmetic alone.

    q153 two-window deconvolution at the IRX2 / IRX2-DT locus with a shared-pair null, GSE252378 · Iqbal Lab oncofusion knowledge graph, beat Q153 · this project 2026
    What this evidence cannot show

    The 19.7% is a one-sided LOWER bound obtained by assuming zero coverage in IRX2-DT's own 215 bp flank; the true attributable fraction could be higher. The complementary '80.3%' is therefore an upper bound on what is shared, not an estimate of it. And the attributable 215 bp is gene-body sequence, not promoter sequence, so even the attributable part does not support a promoter claim. n = 1 sample.

    Evidence that argues against this (1)
    • p454 never cleared the gate (1.898 against 2.0 in the second antibody), so no positive claim was ever made and nothing has to be withdrawn. The entry records a caveat on a sentence that does not yet exist, which is the correct time to record it but also the reason it should not be read as a correction to q151. Iqbal Lab oncofusion knowledge graph et al. 2026
Connected to