Three of eight CUT&RUN libraries are too shallow for a zero to mean anything — measured, not assumed
Finding nothing at a gene only means something if the experiment was deep enough to have found it. In three of these eight samples it was not, and we say so instead of counting them as negatives.
The same thing, for a scientist
A missing value is not a negative value, and this beat needed a number that says which. For each library the fraction of the 19,404 promoters with **zero** coverage was computed; the chance that a promoter with no real signal reads zero in **both** KMT2A-N antibodies is the product. It ranges from **0.019 in p279 to 0.260 in TB986292** — a 14-fold spread [E1]. Zeros in p279, p318, p454, p179 and p247 (P = 0.019-0.061) are informative; zeros in TB051939, TB920488 and TB986292 (P = 0.135-0.260) are not, and every IRX1 absence in those three is reported as **weakly measured** rather than as an absence.
The measurement
| patient | zero-rate IgG | N1 | N2 | P(zero in BOTH N) | verdict |
|---|---|---|---|---|---|
| p279 | 80.1% | 31.5% | 6.0% | 0.019 | informative |
| p454 | 71.7% | 27.7% | 15.3% | 0.042 | informative |
| p179 | 63.7% | 13.9% | 23.3% | 0.032 | informative |
| p247 | 56.7% | 20.1% | 17.0% | 0.034 | informative |
| p318 | 75.7% | 25.8% | 23.5% | 0.061 | informative |
| TB920488 | 80.7% | 24.4% | 55.3% | 0.135 | not informative |
| TB051939 | 74.9% | 34.9% | 40.2% | 0.140 | not informative |
| TB986292 | 94.2% | 30.3% | 85.8% | 0.260 | not informative |
Note the IgG rates: 57-94% of promoters are zero in the IgG controls. So "IgG is zero at my locus" is nearly always true and is close to worthless as a floor on its own — which is why this control, not the IgG comparison, carries the absence argument in this beat.
The separate question the zero-rate does NOT answer: mappability
A locus can also read zero everywhere because the reference cannot be mapped there — these libraries were aligned to a repeat-masked hg38, and IRX1 sits 3.6 Mb from the 5p telomere. The test for that is not the zero-rate but whether the locus is ever non-zero anywhere: 42 of the 66 libraries have non-zero coverage at the same IRX1 window 2. The window is mappable, so a zero in a given sample is a property of that sample.
These are two different failure modes and they need two different controls. The pre-registration separated them (U3 versus A3a) and both were run.
The assumption inside the product, stated because it is load-bearing
Multiplying the two zero-rates assumes the two antibodies fail independently. They are different reagents in separately prepared libraries, so this is defensible, but if library depth is a shared cause the true joint probability is higher than the product. That error runs in one direction only: it makes the shallow libraries even less informative than stated, and leaves the deep-library conclusions where they are. So the assumption is conservative for the claims that depend on it and optimistic for the ones that were withheld anyway.
| n promoters | 19404 |
|---|---|
| irx1 nonzero libraries | 42 of 66 |
p zero both N — the numbers behind the chart
{
"p279": 0.019,
"p318": 0.061,
"p454": 0.042,
"p179": 0.032,
"p247": 0.034,
"TB051939": 0.14,
"TB920488": 0.135,
"TB986292": 0.26
}informative
[ "p279", "p318", "p454", "p179", "p247" ]
not informative for absence
[ "TB051939", "TB920488", "TB986292" ]
igg zero rate range
[ 0.567, 0.942 ]
- An absence scored from a shallow library and written into the graph as a negative would be indistinguishable, in every downstream use, from a measured negative - and the three shallow libraries here are exactly the ones that would have made the IRX1 story look cleaner than it is.
- Should this corpus require a per-library zero-rate beside every CUT&RUN absence it records? Nothing currently does, and absences from this assay family are used across many entries.
- What is the equivalent control for a bigWig-derived MEAN, as opposed to a zero? A low non-zero mean in a shallow library has no analogous bound here.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Promoter zero-rates across the eight KMT2A::AFF1 sample records give P(a promoter reads zero in both KMT2A-N antibodies by chance) ranging from 0.019 to 0.260.
q145 re-analysis of GSE252378: IRX1 and HOXA oncoprotein occupancy across seven KMT2A::AFF1 patients (analyses/q145-irx1-in-hoxa-negative-infants) · Iqbal Lab oncofusion knowledge graph, beat q145 · internal analysis over deposited GSE252378 bigWigs 2026What this evidence cannot show
Computed on one promoter definition (TSS +/- 1 kb, 19,404 canonical TSSs); a different promoter width would give different zero-rates. The independence assumption in the product is stated in the entry and runs conservative.
Evidence that argues against this (1)
- If the two N antibodies' zeros are strongly correlated through depth, the deep-library P values understate the chance of a spurious double zero, and the p179/p247 IRX1 absences - which carry part of the beat's main result - weaken accordingly. Janssens DH et al. 2024
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The IRX1 promoter window has non-zero coverage in 42 of 66 GSE252378 libraries, establishing that it is mappable under the deposit's repeat-masked hg38 bowtie2 pipeline.
GSE252378 - AutoCUT&RUN profiling of MLL-rearranged leukemia samples (hg38 bigWigs and fragment BEDs) · Janssens DH, Ahmad K, Henikoff S · NCBI GEO 2024What this evidence cannot show
Mappability is demonstrated empirically rather than computed from a mappability track; a locus can be partially mappable in a way this test would not reveal.
Evidence that argues against this (1)
- If the 42 non-zero libraries are systematically the deeper ones, mappability and depth are confounded in this test and the argument is weaker than it appears. Janssens DH et al. 2024
the control that decides which IRX1 zeros count as absences
- part of →Plain-English: chromatin and epigenetics — the sticky notes on the DNA, and which ones matter here
a CUT&RUN interpretation rule