The coverage guard withheld 30 of 351 pairs and named every one: 26 contain 152985, which calls zero promoters, and a Jaccard with a zero-call sample is a division, not a measurement
Some comparisons were refused rather than reported, because one of the two samples had no detectable signal at all. Refusing to print a number there, and saying which comparisons were refused and why, is the point.
The same thing, for a scientist
The verdict function refuses to return a Jaccard when any of three pre-registered guards fires: universe below 1,500 promoters, comparable fraction below 10% of 19,404, or either sample calling fewer than 30 promoters. In the primary cell it withheld **30 of 351 pairs**, all three times on the call-count guard. 26 of the 30 contain **152985**, the revumenib-exposed AML that calls zero promoters in every universe; 5 contain A73051 and the rest are singletons. No pair was withheld for insufficient universe: the smallest pair universe is well above the floor. Every withheld pair's coverage figures are recorded beside its refusal, so the refusal is auditable rather than silent.
What was withheld and why
| sample appearing in a withheld pair | count |
|---|---|
| 152985 (AML during revumenib, 0 calls) | 26 |
| A73051 | 5 |
| A82485, TB000694, TB986292, TB995291 | 2 each |
| 148752, A70498 | 1 each |
All 30 refusals were on call count below 30. Zero refusals on universe size and zero on comparable fraction - the 27 pair universes range from 4,646-equivalent upward and every one clears 1,500 promoters and 10% comparability by a wide margin.
Why a zero-call sample cannot contribute an overlap
If one set is empty, the intersection is zero and the union is the other set, so the Jaccard is exactly 0 - a number with the shape of a measurement and the content of an arithmetic identity. Worse, including such pairs in a cohort median pulls the median down and makes every other pair look enriched by comparison. q144's cohort median of 0.132 over 351 pairs includes them. This beat's comparator distributions exclude them and say so.
A missing value is not a negative value. 152985 has no detectable oncoprotein, which is the paper's own finding and a real biological result about menin inhibition. It is not a sample in which the oncoprotein targets zero promoters; it is a sample in which the oncoprotein is not measurable.
The guard is inside the function, not in anybody's head
scored_pair() computes the coverage dictionary first, tests the three guards, and returns a Withheld object carrying the reasons and the coverage when any fires. There is no code path that returns a Jaccard without having passed them, and no printed rate anywhere in this beat appears without its universe size and comparable fraction printed above it.
| script | analyses/q146-one-universe-jaccard/run146.py |
|---|---|
| status | run |
| result summary | 30 of 351 pairs withheld in the primary cell, all on the <30-calls guard; 26 contain 152985 (zero calls). No pair failed the universe or comparable-fraction guards. Every refusal carries its coverage figures. |
| n withheld | 30 |
| n pairs | 351 |
| withheld containing 152985 | 26 |
| n withheld on universe | 0 |
| n withheld on comparable fraction | 0 |
inputs
[ "analyses/q146-one-universe-jaccard/results.json" ]
guards — the numbers behind the chart
{
"min_universe": 1500,
"min_comparable_fraction": 0.1,
"min_calls": 30
}- A cohort median that includes drug-silenced samples understates the chance floor for every other pair, which makes any therapeutic-target overlap look more reproducible than it is.
- The guard fires with one message for two causes - a drug-silenced sample and a shallow library. Should the call-count guard be split so a reader can tell 'no oncoprotein' from 'not enough reads'?
- How many of this corpus's existing cross-sample overlap figures include a zero-call or near-zero-call sample in their denominator?
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
152985's absence of detectable oncoprotein signal is the paper's own most specific claim about this pair: menin-inhibitor treatment reduced the genome-wide MLL N- and C-terminal signal to undetectable levels.
MLL oncoprotein levels influence leukemia lineage identities · Janssens DH, Duran M, Otto DJ, Wu W, Xu Y, Kirkey D, Mullighan CG, Yi JS, Meshinchi S, Sarthy JF, Ahmad K, Henikoff S · Nature Communications 15:9341 2024What this evidence cannot show
Undetectable under the paper's own thresholding with its own calibrator; a different threshold could in principle recover a small number of sites, and this beat did not test how far the threshold would have to move.
Evidence that argues against this (1)
- ENL persists on numerous oncoprotein-target loci in that same sample, per the paper's abstract - so the locus set is not gone, only the oncoprotein channel's ability to see it, which is a reason to treat the zero as an instrument limit and not as an absence of biology. Janssens DH et al. 2024
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
The deposited data for 152985 and for every withheld pair are the GSE252378 hg38 fragment BEDs; the coverage figures recorded beside each refusal are computed from them.
GSE252378: AutoCUT&RUN of KMT2A-rearranged primary leukemia samples (KMT2A N- and C-terminal antibodies, ENL, DOT1L, H3K4me3, IgG) · Janssens DH, et al. · NCBI Gene Expression Omnibus 2024What this evidence cannot show
Coverage here means fragments in a TSS+/-1kb window; a promoter with adequate coverage in this sense can still be uninformative if its fragments are non-specific background.
Evidence that argues against this (1)
- A70498 and A73051, two of the other withheld samples, are low-depth libraries rather than drug-exposed ones, so the guard is firing on two different causes with one message - depth and biology - and the message does not distinguish them. Janssens DH et al. 2024
the guard that produced the headline's coverage-first reporting
the zero-call sample is the menin-inhibited one and is withheld rather than scored as zero overlap
zero-call pairs in the 351-pair median are part of why 0.069 is so low