When a 2 kb promoter call fires at a window shared with another gene, the neighbour fires too 91.3% of the time - against 4.3% for random promoter pairs
also called promoter call discriminability · shared window occupancy agreement
If two genes' measurement windows overlap, and the test says one of them is switched on, it says the other one is switched on as well nine times out of ten. For two unrelated genes picked at random the same test almost never agrees. So the failure is caused by the overlap, not by the test being vague.
The same thing, for a scientist
Generalisation of the IRX2 question over every overlapping pair in this project's window table. There are **1,712** pairs of rows whose 2 kb windows overlap at all and **1,563** that share at least half their sequence. The q147 occupancy rule was applied to both members of each pair in all 27 GSE252378 samples. For the **≥50%-sharing** pairs, of **9,315** pair × sample comparisons in which **at least one** member is called occupied, the two calls **disagree in 809 = 8.7%**, i.e. **they agree 91.3% of the time** [E1]. Over all overlapping pairs the disagreement is 12.8%. **The null matters more than the number.** The same rule, the same libraries and the same 1,563-pair count, but pairing promoters at random: of 8,583 comparisons containing a positive, the calls disagree in **8,217 = 95.7%** [E1]. **The rule discriminates two arbitrary promoters almost perfectly and two overlapping promoters almost not at all** — a 95.7% to 8.7% collapse that is a property of the windows, not of the threshold. Consequences, stated exactly: **8,506 of the 42,201 ≥50%-sharing comparisons are double positives**, each of which is one occupancy statement that two genes have equal claim to; and only **366 of 1,563 pairs (23.4%)** ever disagree in any of the 27 samples, so for three quarters of shared-window pairs the measurement never once separates them [E1].
The table
Rule, from q147 verbatim: OCCUPIED = within-sample share ≥ 2.0 in both KMT2A-N antibodies and above that sample's 95th-percentile floor over eight loci that cannot be occupied in a leukaemia (ALB, RHO, MYH7, MYOD1, NEUROD1, KRT14, GFAP, INS).
| pair set | n pairs | comparisons | both occupied | both not | disagree | disagree, of comparisons with ≥1 positive |
|---|---|---|---|---|---|---|
| overlap at all (<2 kb apart) | 1,712 | 46,224 | 9,008 | 35,893 | 1,323 | 12.8% |
| share ≥50% (<1 kb apart) | 1,563 | 42,201 | 8,506 | 32,886 | 809 | 8.7% |
| NULL: random pairs | 1,563 | 42,201 | 366 | 33,618 | 8,217 | 95.7% |
The null row is the one to read twice. Random pairs almost never co-occur (366 double positives against 8,506), and when one of them fires the other essentially never does. That is what a working discriminator looks like. The overlapping rows are the same instrument on windows that are mostly the same DNA.
What this does NOT say
It does not say the q147 rule is wrong, or that its positive calls are false. A double positive is a true statement that this 2 kb of chromosome 5 (or 19, or 4) carries fusion signal. Everything in that sentence is supported. What is not supported is the gene name attached to it.
It also does not say the shared windows carry no independent information at all. They carry some: 809 disagreements exist, 366 pairs disagree at least once, and PAF1 versus MED29 differ by 0.73 share units in one sample without ever changing a call. The unique flanks are real sequence with real coverage. They are simply too small a fraction of a 2 kb mean to move a threshold, which is quantified in ana-q153-two-window-deconvolution-bounds-what-is-attributable-at-a-shared-promoter.
And it does not generalise beyond promoter-scale means in this deposit. A base-resolution summit call, a narrow-peak caller, or a gene-body window would each behave differently, and none of them has been run here.
| script | analyses/q153-can-an-irx2-promoter-measurement-be-attributed/q153_discriminate.py |
|---|---|
| status | run |
| result summary | A 2 kb promoter occupancy call agrees with its overlapping neighbour 91.3% of the time and with a random promoter 4.3% of the time. The instrument discriminates; the windows do not. |
| n overlapping pairs in table | 1712 |
| n pairs sharing ge50pct | 1563 |
inputs
[ "C:\\Temp\\q147part\\promoter_matrix.tsv (19,404 x 133, RAW mean coverage)", "C:\\Temp\\q145irx1\\promoters.tsv", "q147 samples147.py" ]
measurements — the numbers behind the chart
[
{
"quantity": "call_disagreement_rate",
"subject": "occupancy call at a promoter versus at a neighbour whose 2 kb window shares >=50% of its sequence",
"value": 8.7,
"units": "percent of comparisons containing at least one positive call",
"n": 9315,
"universe": "1,563 pairs x 27 GSE252378 samples = 42,201 comparisons, of which 9,315 have at least one member called OCCUPIED under the q147 rule",
"null": 95.7,
"note": "null is the same statistic over 1,563 RANDOM promoter pairs (8,217 of 8,583), same rule and same libraries, seed 20260924, unmatched"
},
{
"quantity": "call_disagreement_rate",
"subject": "same, over all pairs whose windows overlap at all (<2 kb apart)",
"value": 12.8,
"units": "percent of comparisons containing at least one positive call",
"n": 10331,
"universe": "1,712 pairs x 27 samples = 46,224 comparisons, 10,331 with at least one positive",
"null": 95.7,
"note": "rises from 8.7% because the set includes pairs sharing as little as 1 bp"
},
{
"quantity": "n_double_positive_calls",
"subject": "pair x sample comparisons where BOTH members of a >=50%-sharing pair are called occupied",
"value": 8506,
"units": "comparisons",
"n": 42201,
"universe": "1,563 >=50%-sharing pairs x 27 samples",
"null": 366,
"note": "null is the double-positive count for 1,563 random pairs over the identical comparisons; each double positive is one occupancy statement two genes have equal claim to"
},
{
"quantity": "fraction_of_pairs_that_ever_disagree",
"subject": ">=50%-sharing promoter pairs, across all 27 samples",
"value": 23.4,
"units": "percent",
"n": 1563,
"universe": "366 of 1,563 pairs disagree in at least one of 27 samples; the other 1,197 never separate",
"null": 54.5,
"note": "null is 852 of 1,563 random pairs disagreeing at least once"
}
]- A target list produced by a promoter-scale occupancy screen contains, at this project's own base rate, about one gene in five whose call is shared with a neighbour, and for three quarters of those the data will never distinguish the two. A programme that validates such a hit by knocking down the named gene and seeing no phenotype may be knocking down the wrong one and discarding a real target.
- Would a matched null - pairing each shared-window promoter with an unrelated promoter of the same baseline occupancy frequency - still give a contrast this large? The unmatched null inflates it and the size of that inflation is unmeasured.
- Which of the 8,506 double-positive calls sit at genes this graph has entries for? The pair identities are persisted, but the cross-tab against the corpus has not been made, and it is the number that says how much of this graph is exposed rather than how much of the genome is.
- Does the same collapse happen for the gene-body-scale rule q147 used for IRX1? Gene bodies of divergent pairs are disjoint, so it should not - but tandem and readthrough pairs have overlapping bodies and that case has not been censused at all.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Applying the q147 two-antibody occupancy rule to both members of every pair of promoter rows whose 2 kb windows share at least half their sequence, across 27 GSE252378 samples, the two calls disagree in 809 of 9,315 comparisons containing at least one positive (8.7%), against 8,217 of 8,583 (95.7%) for 1,563 randomly paired promoters under the identical rule and libraries.
q153 promoter-call discriminability over overlapping window pairs, GSE252378 · Iqbal Lab oncofusion knowledge graph, beat Q153 · this project 2026What this evidence cannot show
The rule is a faithful but not byte-identical reimplementation of q147's: it reproduces FLOOR-B and the >=2.0-in-both-N gate and drops the depth-grade UNMEASURED relabelling of zeros, which cannot create a disagreement between two rows and therefore cannot inflate the headline. The matrix was asserted at load to be RAW mean coverage (column means 0.9177-18.7951, 0 of 133 within 1% of 1.0) and divided by column means to give shares, as q147 does. Pairs are counted from the 19,404-row table only, so the true number of overlapping pairs is larger - the table cannot see NR_-only neighbours such as AFF1-AS1. The 27 samples come from 24 patients, so comparisons are not fully independent, and the pair x sample comparisons within a pair are certainly not independent of each other; no attempt is made to compute a p-value from them and none is quoted. The random-pair null is drawn with replacement from the same symbol list at seed 20260924 and is matched on nothing except count and library set - in particular it is NOT matched on expression, CpG content or baseline occupancy rate, so part of the 95.7% is simply that a random promoter is usually unoccupied.
Evidence that argues against this (1)
- The null is unmatched, and that inflates the contrast. A fairer null would pair each shared-window promoter with a DIFFERENT promoter matched on its own occupancy frequency, which would raise the random-pair double-positive count above 366 and lower the 95.7%. That analysis was not run. The direction of the bias is known and the 8.7% side - the measured quantity - is unaffected by it. Iqbal Lab oncofusion knowledge graph et al. 2026
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
CUT&RUN maps protein-DNA binding sites by antibody-targeted MNase cleavage and paired-end sequencing of the released fragments; it is a DNA-fragment readout and nothing in its design assigns a fragment to a transcriptional strand, so a CUT&RUN track cannot separate two transcription start sites 215 bp apart on opposite strands by strand.
An efficient targeted nuclease strategy for high-resolution mapping of DNA binding sites · Skene PJ, Henikoff S · eLife 2017What this evidence cannot show
Abstract read in full via efetch XML on 2026-09-24. `strand-specific` is in expected_absent_terms deliberately: the founding paper does not claim strand specificity because the assay has none, and that ABSENCE is the support for this claim. The inference that a non-strand-specific assay cannot resolve two TSSs by strand is the author's, not the paper's - the paper is cited for what the assay measures, not for this conclusion. Note also the other half of the paper's claim, which cuts the OTHER way: CUT&RUN is explicitly a HIGH-RESOLUTION method, so base-resolution coverage from these same libraries could in principle place a summit within 215 bp. The limitation here is the 2 kb window, not the assay.
Evidence that argues against this (1)
- Because CUT&RUN is high-resolution by design, this beat's conclusion is a criticism of the 2 kb window convention rather than of the deposit. The bigWigs that produced this matrix retain base-resolution information that the promoter mean throws away, so the attribution failure documented here is recoverable from the same raw data - a point that argues for re-analysis rather than for new experiments. Skene PJ et al. 2017
the single worked pair behind the aggregate 91.3% agreement, and the case where the rule actually fires rather than agreeing on zeros
measures the concept: 8.7% disagreement at shared windows against 95.7% for random pairs under the identical rule
same rule, a second limitation: the rule that passes a median 10.75% of promoters also cannot separate two overlapping ones