43.9% of every occupancy call this project's rule makes is at a shared window it cannot attribute - twice the 21.6% base rate, in 27 of 27 samples
also called shared window enrichment among occupied promoters · occupancy call attributability
One promoter window in five in this project overlaps another gene's. But among the windows the fusion-occupancy test actually calls positive, it is nearly one in two. So the calls the test makes are twice as likely as chance to belong to two genes at once.
The same thing, for a scientist
The census says **21.60%** of this project's 19,404 promoter windows contain another gene's TSS within 1 kb. The calls are not drawn uniformly from that set. Applying the q147 occupancy rule to **all 19,404 rows** in each of the 27 GSE252378 samples produces **53,180 occupied calls**, of which **23,331 = 43.87% are at shared windows** — a **2.03-fold** enrichment over the base rate [E1]. It holds in **27 of 27 samples** with no exceptions, range **41.3% to 50.8%**, per-sample median 44.0% [E1]. **So very nearly half of every promoter-scale occupancy statement this machinery can produce is a statement two genes have equal claim to**, and combined with the 91.3% call-agreement measured at such pairs, roughly four in ten of those statements will be duplicated verbatim at the neighbour [E2]. **The direction of the explanation matters and the honest reading is a confound.** Divergent, CpG-island, bidirectional promoters are among the most constitutively active promoters in the genome, and the q147 rule is permissive at promoter scale — it calls a median 10.75% of all promoters occupied per sample. An active promoter is more likely to carry signal for reasons that have nothing to do with KMT2A fusions. **That explanation does not reduce the consequence**: whether the enrichment is biology or activity, the calls the rule makes are twice as likely as chance to be unattributable, and no reader of an occupancy list can tell which of its members are in the 43.9% without running this census [E1].
Per-sample, all 27, because a pooled ratio can hide a single driver
| sample | occupied calls | at shared windows | fraction |
|---|---|---|---|
152985 | 421 | 214 | 0.508 |
A93949 | 1,417 | 663 | 0.468 |
A89691 | 2,248 | 1,043 | 0.464 |
148752 | 2,154 | 984 | 0.457 |
A82485 | 2,062 | 942 | 0.457 |
A93692 | 2,045 | 933 | 0.456 |
p318 | 2,611 | 1,164 | 0.446 |
R3595 | 2,413 | 1,073 | 0.445 |
TB971992 | 2,568 | 1,143 | 0.445 |
TB986292 | 930 | 413 | 0.444 |
| … 14 samples between 0.421 and 0.440 … | |||
p247 | 2,719 | 1,146 | 0.421 |
TB995291 | 1,687 | 696 | 0.413 |
Every one of the 27 is above 0.216. The tightest reading of that is not the mean but the unanimity: 27 of 27, with the minimum (0.413) nearly twice the base rate. The two samples at the extremes (152985 at 0.508, TB995291 at 0.413) are also two of the shallowest, so the spread is consistent with sampling noise on small call counts rather than with a biological gradient.
Why this is a confound and why that does not help
The likely cause is promoter activity, not fusion preference. A head-to-head divergent pair typically shares one CpG-island bidirectional promoter, and those are the promoters most likely to be nucleosome-depleted, marked and transcribed in any cell. The q147 rule scores within-sample share of coverage relative to the average promoter, so any generically active promoter is more likely to clear 2.0. Under that explanation the 2.03-fold is a property of chromatin, and this project's rule is simply finding active promoters.
That reading is not reassuring, for two reasons.
- The consequence is unchanged. However the enrichment arose, 43.9% of the calls are at windows where the gene name is not determined by the data. A caveat is needed on the calls, not on the explanation.
- It is an alternative reading of the rule itself. If shared windows are enriched because they are active, then so are the other 56.1%: the rule may be reporting promoter activity more than fusion occupancy across the board. That is a bigger claim than this beat can support, and
finding-q151-the-two-antibody-2x-occupancy-rule-calls-a-tenth-of-all-promoters-occupiedalready records the permissiveness that makes it plausible.
What would separate the two readings is in hyp-q153-bidirectional-promoters-are-over-represented-among-fusion-occupied-windows-because-they-are-constitutively-active.
| script | analyses/q153-can-an-irx2-promoter-measurement-be-attributed/q153_enrichment.json |
|---|---|
| status | run |
| result summary | 43.87% of 53,180 occupancy calls are at shared windows against a 21.60% base rate; 27 of 27 samples, range 0.413-0.508. Likely a promoter-activity confound, and the attribution consequence is unchanged either way. |
measurements
[
{
"quantity": "fraction_of_occupancy_calls_at_shared_windows",
"subject": "all occupancy calls made by the q147 rule over all 19,404 promoters in 27 GSE252378 samples",
"value": 43.87,
"units": "percent",
"n": 53180,
"universe": "53,180 occupied calls over 19,404 promoters x 27 samples; shared = another refGene gene's TSS within 1 kb, i.e. >=50% window sharing",
"null": 21.6,
"note": "2.03-fold enrichment; the null is fixed by genome geometry and not derived from the data. 27 of 27 samples above the null, range 41.3-50.8%"
},
{
"quantity": "n_samples_above_the_base_rate",
"subject": "per-sample fraction of occupancy calls at shared windows",
"value": 27,
"units": "samples",
"n": 27,
"universe": "27 GSE252378 KMT2A-fusion samples; base rate 0.2160; per-sample minimum 0.4126, maximum 0.5083, median 0.4398",
"null": null,
"note": "unanimity is quoted instead of a p-value because the 19,404 within-sample calls are not independent - neighbouring promoters share coverage, which is the thing being measured"
}
]- Any occupancy-derived gene list from this deposit should be read as roughly 44% region-level and 56% gene-level. A programme that prioritises targets by how many samples show occupancy will systematically favour bidirectional CpG-island promoters - which is to say, will systematically favour the windows whose gene assignment is least determined.
- Is the 2.03-fold explained entirely by promoter activity? Matching shared and unshared windows on IgG signal, on CpG-island overlap, or on the per-promoter share in the KMT2A-C arms would test it, and all three are computable from what is already on disk.
- Does the enrichment persist under a stricter rule? At a 4x or 8x share gate the call count falls sharply and the shared fraction may or may not; the sensitivity was not computed.
- How many of the 23,331 shared-window calls are at genes this graph has entries for? That cross-tab is the number that tells the corpus how exposed its own claims are, and it has not been made.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Applying the q147 two-antibody occupancy rule to all 19,404 promoter rows in each of 27 GSE252378 samples yields 53,180 occupied calls, of which 23,331 (43.87%) are at windows containing another gene's TSS within 1 kb, against a base rate of 21.60%; the enrichment holds in 27 of 27 samples with a range of 41.3% to 50.8%.
q153 enrichment of shared-window promoters among occupancy calls, GSE252378 · Iqbal Lab oncofusion knowledge graph, beat Q153 · this project 2026What this evidence cannot show
The occupancy rule is a faithful reimplementation of q147's (share >= 2.0 in both KMT2A-N libraries AND above the 95th-percentile floor over ALB, RHO, MYH7, MYOD1, NEUROD1, KRT14, GFAP, INS) without its depth-grade UNMEASURED relabelling of zeros, which can only affect zero rows and so cannot create a positive call. The input matrix was asserted at load to be RAW mean coverage (column means span 0.9177-18.7951, 0 of 133 within 1% of 1.0) and converted to shares by dividing by the column mean. The base rate 21.60% is fixed by genome geometry and is not derived from the data, so this is a genuine enrichment test and not a self-stratified ratio. No p-value is quoted: the 19,404 calls within a sample are not independent (neighbouring promoters share coverage by construction, which is the very thing being measured), so any nominal p would be fabricated precision. The unanimity - 27 of 27 with a minimum of 0.413 - is offered instead. 27 samples from 24 patients.
Evidence that argues against this (1)
- The most likely cause is that divergent CpG-island promoters are constitutively active and therefore more likely to clear a within-sample-share threshold for reasons unrelated to KMT2A fusions. If so, the 2.03-fold says nothing about fusion biology - it says the rule finds active promoters. The enrichment was NOT matched on promoter activity, CpG content, or IgG or input signal, and until it is, the confound explanation is at least as well supported as any other. Trinklein ND et al. 2004
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Of the occupancy calls at shared windows, the identical call is made at the overlapping neighbour 91.3% of the time, so roughly four in ten of all occupancy statements this rule produces are duplicated at a second gene.
q153 promoter-call discriminability over overlapping window pairs, GSE252378 · Iqbal Lab oncofusion knowledge graph, beat Q153 · this project 2026What this evidence cannot show
The 91.3% is measured over pairs BOTH of whose members are rows of the 19,404-row table, which is a subset of shared windows - where the neighbour is NR_-only (as for AFF1-AS1) or a microRNA (as for CDK9) the neighbour has no row and no call, so the duplication cannot be observed even though the sequence is still shared. 'Roughly four in ten' multiplies 43.87% by 91.3% and is an approximation stated as one.
Evidence that argues against this (1)
- A duplicated call is not a false call. Both statements are true of the DNA; only the gene attribution is undetermined. A reader who cares about which regions the fusion occupies loses nothing here - the loss falls entirely on anyone reading the list as a list of genes. Iqbal Lab oncofusion knowledge graph et al. 2026
runs the C-arm control that entry named as its cheapest gap; N minus C is -1.00 points
reproduces q153s sharing census on the current annotation, 14.19% and 22.63%
the 21.60% base rate this 43.87% is measured against
the same rule's permissiveness is what makes the promoter-activity explanation of the 2.03-fold plausible
the 2.03-fold enrichment this hypothesis exists to explain