The same occupancy rule passes 0.61% of gene bodies and 10.75% of promoters, so IRX1 at 3 of 27 is 18x its gene-scale expectation where at promoter scale it was below it
also called q154 T1 · the gene-scale pass rate · the missing denominator for q147
The test this project uses to decide whether the fusion protein is sitting on a gene passes about one promoter in ten - so any promoter looks occupied in two or three of the twenty-seven patients just by chance. Applied to whole gene bodies instead, the same test passes only about one gene in a hundred and sixty, so chance gives you a sixth of one patient. That is why the IRX1 result looks weak at one scale and strong at the other: the two numbers were never comparable.
The same thing, for a scientist
The occupancy rule - within-sample share >= 2.0x an average promoter in **both** KMT2A-N libraries, above that library's cannot-be-occupied floor - had never been calibrated at gene-body scale, because no gene-body matrix existed. Built here over **all 19,402 hg38 refGene gene bodies x 133 GSE252378 libraries**, the rule passes a **median 0.608% of gene bodies per sample** (min 0.175% in 152985, the drug-exposed library; max 0.974% in p318), against **10.75% of promoters** - **17.7-fold more stringent** [E1]. **532 of 19,402 gene bodies (2.74%) are occupied in at least one of the 27 samples**, against 6,049 of 19,404 promoters (31.2%) [E1][E2]. The consequence for the project's headline is direct. By linearity of expectation an arbitrary gene body is occupied in **0.164 of 27 samples**; an arbitrary promoter is occupied in **2.90 of 27**. **IRX1 is occupied in 3 of 27 at gene-body scale - 18.3x its expectation - and in 2 of 27 at promoter scale, 0.69x its expectation** [E1]. The scale disagreement that three cycles read as a contradiction is a difference in denominator, not in biology. **The pass rate is identical under the flat and the length-matched floor** (median 0.608% in both), because the floor never binds at this scale, and the six-member AFF1-exclusive class is the same six genes under all three floor definitions [E1].
The two scales side by side
| promoter scale (2 kb, 19,404) | gene-body scale (19,402 full gene bodies) | |
|---|---|---|
| median pass rate per sample | 10.75% | 0.608% |
| min / max per sample | 1.86% (A80956) / 14.01% (p247) | 0.175% (152985) / 0.974% (p318) |
| occupied in >= 1 of 27 | 6,049 (31.2%) | 532 (2.74%) |
| expected samples occupied, arbitrary feature | 2.90 of 27 | 0.164 of 27 |
| IRX1 | 2 of 27 = 0.69x expectation | 3 of 27 = 18.3x expectation |
| AFF1-exclusive at k>=1 vs its permutation null | 332 vs 425 (below) | 63 vs 74 (below) |
| AFF1-exclusive at k>=3 vs its permutation null | 17 vs 9.1 (p 0.112) | 6 vs 0.88 (p 0.0026) |
The k>=1 row is the one that reproduces across scales: a singleton AFF1-only feature is at or below chance at both scales. What changes is the recurrent class - at gene-body scale a feature occupied in three or more AFF1 samples and no comparator is 6.8 times rarer under permutation than observed, where at promoter scale it was not distinguishable from chance 12.
Why a gene body is a harder gate than a promoter, stated as the mechanism
The denominator in both cases is the mean coverage of an average promoter in that library - that is q147's unit and it was kept unchanged so the 2.0x gate means the same thing at both scales. A CUT&RUN for a promoter-binding factor puts most of its signal in promoters, so the mean gene body sits well below the mean promoter: measured here, the mean gene-body share per library spans 0.140 to 0.987 across all 133 libraries, never reaching 1.0 1. Asking a gene body to reach 2.0x an average promoter is therefore asking it to be 2.0 to 14 times its own scale's mean, where a promoter must only be twice its own.
This is not a defect of the rule; it is the reason the gene-body call is worth more. But it means the two pass rates are the only honest way to compare a gene-body call with a promoter call, and until this beat neither the rule's authors nor its critics had one.
The instrument gate, which had to pass before any of this counted
The matrix is built from streamed fragment midpoints, not from bigWigs, so the pre-registration required it to reproduce q147's bigWig-derived values before its null could be applied to q147's calls. Over the 21 named q147 regions x the 79 libraries q147 scored the two agree at Spearman 0.9901 over 1,659 region x library pairs, with a per-region median ratio of 1.035 (range 0.800-1.177); at IRX1's gene body the ratio is 0.973 and at the HOXA9/HOXA10 20 kb region 1.010 1. Scored with q147's own regions and q147's own 95th-percentile floor, this instrument returns IRX1 OCCUPIED in exactly p279, p318, p454; 0 of 18 non-AFF1; HOXA occupied in 17 of 18 - q147's table to the sample 1.
| script | analyses/q154-gene-scale-occupancy-null/analyze.py |
|---|---|
| status | run |
| result summary | Gene-body pass rate median 0.608% of 19,402 per sample vs 10.75% of 19,404 promoters; 532 occupied anywhere vs 6,049; IRX1 is 18.3x its gene-scale expectation and 0.69x its promoter-scale one. |
| n gene bodies | 19402 |
| n libraries | 133 |
| genome build | hg38 |
| units | interval fragment-midpoint density / that library's own mean promoter density; 1.0 = an average promoter |
inputs
[ "GSE252378 GSM*_hg38.bed.gz, all 133 libraries, streamed", "C:/Temp/q145irx1/promoters.tsv and refGene.txt.gz (hg38)", "C:/Temp/q144mag/counts (per-library promoter fragment counts, the denominator)" ]
measurements — the numbers behind the chart
[
{
"quantity": "occupancy_rule_pass_rate",
"subject": "fraction of gene bodies called occupied per sample by the two-antibody 2.0x-plus-floor rule",
"value": 0.608,
"units": "% of gene bodies per sample (median over 27 samples)",
"n": 19402,
"universe": "19,402 hg38 refGene gene bodies x 27 GSE252378 KMT2A-fusion samples, KMT2A-N1 and N2, shares against each library's own mean promoter density",
"null": 10.75,
"note": "the null quoted is the same rule's promoter-scale pass rate; min 0.175% (152985), max 0.974% (p318)"
},
{
"quantity": "expected_samples_occupied_for_an_arbitrary_feature",
"subject": "an arbitrary gene body, by linearity of expectation over 27 samples",
"value": 0.164,
"units": "samples of 27",
"n": 27,
"universe": "19,402 gene bodies x 27 samples at a 0.608% median pass rate",
"null": 2.9,
"note": "the null quoted is the promoter-scale equivalent; IRX1 at 3 of 27 is 18.3x the gene-scale figure and 0.69x the promoter-scale one"
},
{
"quantity": "n_features_occupied_in_at_least_one_sample",
"subject": "gene bodies occupied in >=1 of the 27 GSE252378 samples",
"value": 532,
"units": "gene bodies",
"n": 19402,
"universe": "the same matrix, flat floor",
"null": 6049,
"note": "the null quoted is the promoter-scale count of 6,049 of 19,404 (31.2%); gene-body is 2.74%"
},
{
"quantity": "instrument_agreement_with_q147",
"subject": "streamed fragment-midpoint shares versus q147's bigWig mean-coverage shares over its 21 named regions",
"value": 0.9901,
"units": "Spearman rho",
"n": 1659,
"universe": "21 regions x 79 KMT2A-N and KMT2A-C libraries; per-region median ratio 1.035 (0.800-1.177), IRX1_gene 0.973, HOXA9_HOXA10_20kb 1.010",
"null": 0.9,
"note": "the null quoted is the pre-registered gate; the gate passed"
}
]- A rule that calls 2,086 promoters per sample occupied will nominate hundreds of therapeutic targets that are simply active promoters. Any target list this project circulates from the promoter-scale rule should carry the pass rate beside it, because a reader who does not know the denominator cannot tell a target from a background call.
- What is the pass rate in a non-leukaemic control? GSE159608 holds CD34+ HSPC libraries and the q144 beat already extracted promoter counts for them; a gene-body pass there would give the first not-a-leukaemia baseline this rule has ever had.
- At what threshold does the gene-body rule's per-sample call count match the authors' ~125 target genes per sample? Setting the two instruments to the same call volume is the only way to compare their specificity rather than their permissiveness.
- The pass rate varies 5.6-fold across samples (0.175% to 0.974%). Does that track sequencing depth, in-promoter enrichment fraction, or oncoprotein level - and if it tracks depth, every cross-sample count in this project inherits it?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
A genome-wide gene-body occupancy matrix over all 19,402 hg38 refGene gene bodies and all 133 GSE252378 libraries gives a median per-sample pass rate of 0.608% (min 0.175%, max 0.974%), 532 gene bodies occupied in at least one of 27 samples, mean gene-body share per library spanning 0.140-0.987, and reproduces q147's gene-scale call table exactly (IRX1 in p279/p318/p454, 0 of 18 non-AFF1, HOXA 17 of 18) with Spearman 0.9901 against q147's bigWig values over 21 regions x 79 libraries.
Q154 genome-wide gene-body occupancy matrix and calibration for GSE252378 · Iqbal Lab oncofusion knowledge graph, beat Q154 · this project 2026What this evidence cannot show
Gene bodies are the txStart-txEnd span of the exact RefSeq accession this project's promoter table uses for each symbol, so a gene with a dominant alternative first exon is scored on one isoform. Two of 19,404 symbols were dropped for gene bodies under 200 bp (ETDC, KRTAP6-2). Fragment-midpoint counting understates short intervals relative to long ones, which biases gene-body shares UPWARD relative to the 2 kb promoter denominator and therefore makes the pass rate and its null conservative in the direction of the rule looking less stringent than it is.
Evidence that argues against this (1)
- A pass rate is a property of one deposit's libraries as well as of the rule. GSE252378's libraries range from 0.75 M to 17.8 M fragments and there is no spike-in, so a rate computed here does not transfer to another CUT&RUN study, and the min of 0.175% belongs to the one library in which a menin inhibitor had evicted the oncoprotein. NCBI Gene Expression Omnibus 2024
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
At promoter scale the identical rule passes a median 10.75% of 19,404 promoters per sample (min 1.86% A80956, max 14.01% p247), leaves 6,049 promoters occupied somewhere, and gives 332 AFF1-only promoters against a label-permutation null of 425.4.
Q152 results - the promoter occupancy matrix, pass rates and the AFF1-only null · Iqbal Lab oncofusion knowledge graph, beat Q152 · this project 2026What this evidence cannot show
q152's promoter numbers include both menin-inhibitor-labelled samples in the 27; this beat's gene-body numbers also cover all 27 for the pass rate and drop only 152985 for the AFF1-exclusive tests. The pass-rate comparison is therefore like-for-like and the null comparison is not exactly so.
Evidence that argues against this (1)
- 10.75% of promoters is not obviously wrong as a pass rate for a promoter-binding oncoprotein complex in a blast, so the comparison should not be read as 'the promoter rule is broken'. What it shows is that the two rates differ by 17.7-fold and that a count of samples means something different at each scale. Iqbal Lab oncofusion knowledge graph et al. 2026
48.3% of those calls are histone-cluster genes, which are 0.5% of the table
supplies the missing gene-body half of the same calibration: 0.608% against that entry's 10.75%
- supports →IRX1 fusion occupancy is confined to KMT2A::AFF1: 0 of 18 samples across five other partners
the gene-body denominator that entry never had: an arbitrary gene body is expected occupied in 0.164 of 27 samples, so IRX1's 3 of 27 is 18.3x expectation
the streaming extractor that made the matrix affordable
the pass rate is the null's denominator
why the pass rate is identical under the flat and the length-matched floor
same gene-body matrix and same rule