The separation MARGIN, not the recurrence count, is the statistic that identifies a partner-specific promoter
also called q152 margin hypothesis
Counting how many patients a gene switch is 'on' in turns out not to distinguish a real partner-specific gene from noise. A different number might: how far above the comparison patients the signal sits. In this dataset the three genes with the biggest gap include the one gene with independent published evidence, which is suggestive and needs a proper test.
The same thing, for a scientist
Q152 showed that the count of promoters occupied in >=k of an index arm and 0 of a comparator arm is not distinguishable from a label permutation (p = 0.11 at k = 3). The hypothesis is that the **margin** - the weakest index-arm occupied value divided by the strongest comparator value - carries the signal the count does not. The suggestive observation is that the three core members with margin > 1.5 are WT1 (2.17), EPHA7 (1.84) and SPRED1 (1.59), and EPHA7 is the only member of the 17 with an independent published KMT2A::AFF1 promoter ChIP (PMID 17726105). That is a post-hoc pattern in a set of seventeen and is explicitly NOT offered as evidence; it is the reason to run the test.
Why the count fails and the margin might not
A count is a function of a threshold crossing, so it inherits the threshold's arbitrariness: Q152 measured that 9 of 17 core promoters have a comparator sample within 10% of the 2.0x gate and one has a comparator sample stronger than its own weakest index call. A margin is continuous and dimensionless and does not require a threshold to be meaningful. It is also directly interpretable: margin 1.0 means the arms are indistinguishable.
The margin has its own failure mode, which must be pre-registered against: it rewards promoters whose comparator readings are near zero, and a comparator reading is near zero both when the promoter is genuinely absent and when the library is shallow. Any margin statistic must therefore be depth-controlled, and this project has already established that there is no coverage-invariant measure in this deposit.
| statement | For a partner-exclusive promoter set, the separation margin (weakest index-arm occupied value / strongest comparator value) discriminates promoters with independent literature support from threshold artefacts, whereas the recurrence count does not. |
|---|---|
| origin | data-pattern |
| prior probability | 0.25 |
predictions
[ "Ranking all 6,049 occupied-anywhere promoters by margin against the AFF1 arm will place a higher proportion of published KMT2A::AFF1 target genes in the top 100 than a recurrence ranking does.", "A label permutation of the margin statistic will show a right tail that the recurrence count does not have, i.e. the observed top-margin promoters will exceed the permutation 95th percentile even though the count does not.", "Margin will correlate positively with comparator-arm library depth, and the effect must be removed before any verdict." ]
analyses
[ "ana-q152-planned-margin-ranked-partner-specificity-scan" ]
falsifiers
[ "If a label permutation of the margin statistic over the 6,049 occupied-anywhere promoters puts the observed maximum margin (WT1, 2.17) at or below the permutation 95th percentile, the margin carries no more signal than the count and the hypothesis dies.", "If the top-100 margin-ranked promoters against the AFF1 arm contain no more of the 40-50 published oncoprotein target genes that q143 recovered than the top-100 recurrence-ranked promoters do, the hypothesis dies.", "If margin correlates with comparator-arm mean library coverage at Spearman > 0.5 across the 6,049 promoters, the statistic is a depth readout and the hypothesis is unusable in this deposit regardless of the first two tests." ]
- If a margin ranking were adopted and it is in fact a depth readout, every partner-specificity claim built on it would be a statement about which libraries were sequenced deepest - the same failure this project already recorded for the onco/neg ratio, which was built to cancel library scaling and still tracks the enrichment axis at +0.631.
- Is there a published partner-specific promoter set for any KMT2A fusion against which a margin ranking could be validated externally, rather than against this project's own q143 recovery list?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
In the 17-promoter AFF1-only core, the separation margin ranges 0.75 to 2.17 with a median of 1.14; only WT1 (2.17), EPHA7 (1.84) and SPRED1 (1.59) exceed 1.5, and EPHA7 is the only one of the 17 with an independent published KMT2A::AFF1 promoter-occupancy experiment.
Q152 per-core-gene separation margins over GSE252378 · Iqbal Lab oncofusion knowledge graph, beat Q152 · this project 2026What this evidence cannot show
Three of seventeen exceeding 1.5, one of which has literature support, is a pattern in a set of seventeen with no null attached. It is a reason to test, not a result.
Evidence that argues against this (1)
- Three of seventeen above a threshold chosen after seeing the data, one of which has literature support, is exactly the shape of a post-hoc pattern. If the margins had been 1.4, 1.5 and 1.6 a different cut would have been chosen; nothing in this observation constrains the hypothesis. Iqbal Lab oncofusion knowledge graph et al. 2026
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
ChIP analysis demonstrates occupancy of tagged ALL1 fusion proteins on the EphA7 promoter and siRNA suppression of ALL1/AF4 in SEMK2 t(4;11) cells down-regulates EphA7.
ALL1 fusion proteins induce deregulation of EphA7 and ERK phosphorylation in human acute leukemias · Nakanishi H, Nakamura T, Canaani E, Croce CM · Proc Natl Acad Sci U S A 104:14442-7 2007What this evidence cannot show
2007, tagged fusion protein, cell lines. This is the only independent occupancy anchor available for any member of the core, so the margin hypothesis rests on n = 1 positive control.
Evidence that argues against this (1)
- EPHA7 is a single positive control from 2007 using tagged fusion protein in cell lines, and the same paper reports AF9 as a positive where this deposit is negative. A hypothesis validated against one control that itself disagrees with the data in part is very weakly anchored. Nakanishi H et al. 2007
the margin distribution the hypothesis is about
proposed because the recurrence count failed
the pre-committed scan and its three falsifiers
the single published positive control the hypothesis rests on
q148 found a promoter class that did not transfer outside the sample that defined it; the same generalisation risk applies