Oncofusion

start / graph / ana-the-aff1-only-promoter-pattern-is-below-its-own-null-and-the-figure-the-irx1-line-rests-on-is-not-an-enrichment

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.

The 332 AFF1-only promoters are BELOW chance: a label-permutation null over the same occupancy matrix expects 426, so the figure the IRX1 partner-specificity line rests on is not an enrichment - and my own brief told the agent the opposite

also called 332 against a null of 426 · the AFF1-only pattern is below chance · independence over-estimates exclusion by an order of magnitude · q151 measurement null was left as null

A count of genes that look switched on only in the samples carrying one fusion partner was treated as evidence that partner is special. Shuffling which samples count as that partner produces MORE such genes, not fewer, so the count is not evidence of anything.

The same thing, for a scientist

Q151 established the first null for this project's occupancy rule and found **332 of 19,404 promoters (1.71%)** show the pattern the IRX1 story turns on: occupied in at least one KMT2A::AFF1 sample and in none of the 19 non-AFF1 samples. **I re-derived 332 exactly.** The number has been read as a specificity signal. 🔴 **It is below chance.** Holding the 19,404 x 27 occupancy matrix completely fixed and redrawing only *which 8 of the 27 samples count as the index arm*, 10,000 times: | | | |---|---:| | observed AFF1-only promoters | **332** | | permutation null, mean | **426.4** | | null median | 423 | | null 5th-95th percentile | **287 - 579** | | null range | 177 - 755 | | p (fraction of draws at or below observed) | **0.1517** | Q152 ran this first and got mean **425**, 5th-95th **288-580**, p **0.1465**; my independent re-run gives **426.4 / 287-579 / 0.1517** [E1]. **The AFF1 arm produces FEWER exclusive promoters than a random 8-of-27 split does** - meaning AFF1 samples *share* occupancy with non-AFF1 samples slightly more than chance, not less. 🟠 **So the 332 cannot support a partner-specificity claim, and IRX1's membership in it is not evidence.** IRX1 is one of 332 promoters in a set that is not enriched, and it sits in the 2-of-8 tier with 47 others. This does not touch q147's **gene-body-scale** result, which is a different measurement with its own floor - but it removes the promoter-scale pattern as independent support for it. 🔴 **And my dispatch brief told the agent the opposite, in a highlighted instruction.** I wrote: *"an independence calculation is a LOWER bound on the expected recurrence"*, citing this project's own recorded caveat that promoters cluster. **For a pattern requiring EXCLUSION it is an over-estimate by an order of magnitude.** On the median 10.75% pass rate, independence expects **102.7** promoters recurring in >=3 of 8 and absent from all 19, against a permutation mean of **9.1** - an **11.3x** over-estimate; on per-sample rates Q152 computes 150.9 and **16.6x** [E2]. **The reason is the direction of the correlation.** Clustering makes *co-occurrence* more likely than independence predicts - which is the caveat this corpus recorded, and it is correct. But the same clustering makes *being absent from all 19 comparators* much **harder** than independence predicts, because a promoter occupied anywhere tends to be occupied in several samples. **The folklore is right about co-occurrence and I applied it to exclusion.**

Go deeper
What this does and does not reach

Reached: the promoter-scale AFF1-only count is not an enrichment, so neither the 332 nor IRX1's place in it is evidence of partner specificity. Q152's separate finding follows: the 17-gene recurrent core (>=3 of 8 AFF1, 0 of 19 non-AFF1) sits against a permutation null of mean 9.1, SD 6.1, 5th-95th 2-21, giving p = 0.112 - also at chance.

Not reached: q147's headline is gene-body scale with a width-matched cannot-be-occupied floor, and no equivalent genome-wide gene-body null exists yet because the corpus holds no gene-body matrix for all 19,404 genes. That calibration is the outstanding item, and until it is run the gene-scale result stands on its own terms while the promoter-scale support for it is withdrawn.

🟠 One correction to Q152's own arm definition, which it flagged and got half right. It removed two comparator samples as menin-inhibitor treated, 148752 and 152985, on the deposit's treated: true field. This corpus established two weeks of cycles ago that GEO's "treated with the Menin inhibitor" is a CASE-level label and that 148752 is the untreated diagnosis sample - only 152985 is drug-exposed. So the informative comparator arm is 18, not 17, and Q152's recomputed p of 0.2163 rests on dropping one sample too many. The direction is unaffected: the observed count stays below the null either way.

Facts and figures
observed aff1 only promoters332
null mean426.4
null median423
p lower0.1517
n draws10000
seed20260924
recurrent core observed17
recurrent core null mean9.1
recurrent core p upper0.1119
independence expectation median rate102.7
independence over estimate factor median rate11.3
independence over estimate factor per sample q15216.6
brief errormy dispatch brief instructed that an independence calculation is a LOWER bound on expected recurrence; for an EXCLUSION criterion it over-estimates by an order of magnitude
arm definition correctionQ152 dropped both 148752 and 152985 as menin-inhibitor treated on the deposit's case-level label; 148752 is the untreated diagnosis sample, so the comparator arm is 18 rather than 17
null 5th 95th
[
  287,
  579
]
null range
[
  177,
  755
]
q152 reported
{
  "mean": 425,
  "p5_p95": [
    288,
    580
  ],
  "p_lower": 0.1465
}
recurrent core null 5th 95th
[
  2,
  21
]
measurements
[
  {
    "quantity": "promoters occupied in the index arm only, observed against a label-permutation null",
    "value": {
      "observed": 332,
      "null_mean": 426.4,
      "p_lower": 0.1517
    },
    "unit": "promoters",
    "subject": "the 8 KMT2A::AFF1 samples against the 19 non-AFF1 samples of GSE252378",
    "universe": "19,404 refGene promoters; occupancy matrix held fixed, arm membership permuted 10,000 times"
  },
  {
    "quantity": "factor by which an independence calculation over-estimates an exclusion pattern",
    "value": 11.3,
    "unit": "ratio of expectations",
    "subject": "promoters recurring in >=3 of 8 index samples and absent from all 19 comparators",
    "universe": "median 10.75% per-sample pass rate against a permutation mean of 9.1; 16.6x on per-sample rates"
  }
]
Still open
  1. Build a genome-wide GENE-BODY occupancy matrix and calibrate q147's gene-scale IRX1 result the same way. The promoter-scale support is withdrawn; the gene-scale claim has never had a null, and the corpus holds no gene-body matrix for all 19,404 genes.
  2. Run a partner-block permutation, drawing the arm from within a single partner group, to test whether exclusivity is a property of any coherent grouping rather than of AFF1. The current null mixes six partners into every draw.
  3. Why is the observed count below the null at all? If AFF1 samples share occupancy with non-AFF1 samples more than chance, that is itself a statement about the shared HOXA programme and could be measured rather than left as a curiosity.
Every source behind this entry (2)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.

    Holding the 19,404 x 27 promoter occupancy matrix fixed and permuting only which 8 of 27 samples form the index arm over 10,000 draws, the number of promoters occupied in at least one arm sample and no comparator sample has mean 426.4, median 423, 5th-95th percentile 287 to 579 and range 177 to 755, against an observed 332, giving p = 0.1517 in the lower tail.

    What this evidence cannot show

    Coordinator re-run 2026-09-24 (cycle 141) at seed 20260924 from C:/Temp/q147part/promoter_matrix.tsv, with the matrix asserted RAW at load and converted by column mean. A label permutation holds the occupancy pattern fixed and so preserves promoter-promoter correlation, which is the point; it does NOT preserve the AFF1 arm's particular composition, so a draw can place two samples from one patient in the arm exactly as the real arm does with p279/p318. Small numeric differences from Q152 (426.4 against 425) are draw-to-draw variation.

    Evidence that argues against this (2)
    • Being below a permutation null is not evidence of anti-specificity either, and should not be reported as one: p = 0.15 in the lower tail is unremarkable. The correct reading is that 332 is consistent with chance, so it carries no information about partner specificity in either direction.
    • The permuted arm is drawn without respect to partner, institution or lineage, so a random 8 mixes AFF1 with ENL, AFDN, AF9, AF10 and EPS15 samples. If exclusivity is a property of ANY coherent biological grouping rather than of AFF1 specifically, this null is conservative in a way that would hide a real effect. A partner-block permutation would test that and was not run.
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.

    For a pattern requiring recurrence in at least 3 of 8 index samples and absence from all 19 comparators, an independence calculation at the median 10.75 percent per-sample pass rate expects 102.7 of 19,404 promoters, against a label-permutation mean of 9.1 - an over-estimate of 11.3-fold, and 16.6-fold on Q152's per-sample rates.

    What this evidence cannot show

    The 11.3x figure uses a single median pass rate for all 27 samples; Q152's 16.6x uses each sample's own rate, which is the better calculation and gives a larger factor. Either way the direction is the finding and the magnitude is rate-model dependent. Both compare against the same permutation mean of 9.1.

    Evidence that argues against this (1)
    • The general folklore the brief invoked is not wrong: for CO-OCCURRENCE, promoter clustering does make an independence calculation a lower bound on expected overlap, and this corpus has recorded that correctly elsewhere. The error was applying a co-occurrence intuition to an exclusion criterion, where the same correlation pushes the other way. Nothing here licenses discarding the original caveat in its own setting.
Connected to