TB051939 is rank 12 of 26 and owns fewer singletons than its even share; the library that owns 22.4% of singletons destroys no exclusivity at all; and dropping any single comparator moves the exclusive set by a median of 2 genes
also called q175 · the loud library that wasn't · exclusivity sensitivity
Three of the four genes that looked specific to this fusion were seen in only one patient sample, so the obvious worry was that that sample is simply noisy and finds things everywhere. It is not - it is middle of the pack, and produces fewer one-off findings than an average sample. A different sample does produce a fifth of all the one-offs, but removing it changes nothing, because its extra findings are in genes that were never candidates. Removing any single comparison sample shifts the answer by about two genes out of 63.
The same thing, for a scientist
q171 closed the occupancy line and left one loose end: three of its four hits - ANGPT1, FOXL2, FOXL2NB - share an identical call vector, each called in **TB051939 alone**. The natural dismissal is that TB051939 is a noisy library. **Two versions of that dismissal were tested. Both are refuted.** 🔴 **TB051939 is unremarkable.** It makes **122 calls, rank 12 of 26**, against a median of 119 and a mean of 113.3 - **z = +0.30**. The loudest library is p318 at 189. More pointedly, of the **219 genes called in exactly one library, TB051939 owns 6 — 2.7%, BELOW its even share of 3.8%** [E1]. *It is a quiet library for singletons.* "Called only in TB051939" is not a statement about the library. 🔵 **The loudest singleton producer is A80956, at 49 of 219 (22.4%, ~6× even share) — and it is a comparator, which should matter more.** A loud fusion library manufactures false exclusives; a loud *comparator* manufactures false NON-exclusives, disqualifying genes that would otherwise pass. So the question became: how many genes does A80956 wrongly exclude? 🔴 **Zero.** Dropping A80956 creates **0 new exclusives** [E2]. Its 49 singletons are genes called by *nothing else, including no fusion library*, so they fail the '≥1 fusion call' half of the criterion whether or not A80956 is present. **Singleton-loudness and exclusivity-damage are different properties and I conflated them** — the statistic that flagged A80956 was the wrong statistic for the question it was flagged by. 🟢 **The genuine sensitivity result is reassuring.** Against a baseline of **63** genome-wide exclusives, dropping each of the 18 comparators in turn adds a **median of 2.0** and a **maximum of 8** (A73051, 12.7% inflation) [E2]. The worst comparator is not the loud one. **IRX1, ANGPT1, FOXL2 and FOXL2NB have zero comparator calls**, so no single-library drop touches any of them. 🟠 **The decision this licenses on the six owed genes.** ANGPT1, FOXL2 and FOXL2NB **cannot be refused on library-quality grounds** — that explanation was tested twice and failed twice. They remain weak for the reason q171 already established and no other: **they are one observation, not three.** Identical call vector, single library, and FOXL2/FOXL2NB 79 bp apart. CLEC14A, ROMO1 and DUSP6 each carry comparator calls and are not exclusive at all. **Recorded as a refusal on the correct grounds, with the incorrect grounds recorded as tested and rejected** — the distinction matters, because a refusal resting on a false reason is a refusal that will be reversed by the next person who checks.
Why the singleton count was the wrong instrument for the question
A80956 looked alarming on a defensible statistic: 22.4% of all single-library calls, nearly six times an even share, in a comparator arm. The inference — that it is disqualifying genes from exclusivity — does not follow, and the reason is structural.
Exclusivity requires two conditions: at least one fusion call, and zero comparator calls. A80956's excess lives entirely in genes satisfying neither — genes it alone calls. Removing it cannot promote them, because they still have no fusion call. To damage exclusivity a comparator must fire on genes that fusion libraries ALSO call, which is a different and much rarer event, and it is what A73051 does eight times.
The general form: a marginal count cannot answer a question about a conjunction. The right statistic was never 'how many singletons does this library own' but 'on how many fusion-called genes is this library the only comparator'. This corpus has the same error recorded before under a different name — two true marginals do not give you the cross-tab.
What this does not establish
It does not rehabilitate the occupancy line. q171's patient-level null returned p = 0.092 at every threshold, and nothing here changes that. Showing that a specific artefactual explanation is absent is not evidence for the finding — an observation can be both non-artefactual and unremarkable, and this one is.
It does not show the libraries are comparable. Raw depth correlates with call count at only r = +0.391 across 26 libraries, so roughly 85% of the variation in how many genes a library calls is not depth. Something else differs between these libraries and it was not identified here. That is a larger uncontrolled factor than anything this analysis ruled out.
It is a leave-one-out, not a leave-k-out. Dropping any single comparator moves the set by a median of 2, but comparators may fail in correlated groups — by institution, by handling batch — and dropping a correlated pair or triple was not tested. A set robust to every single deletion can still be fragile to the right pair.
| script | analyses/q175-is-tb051939-just-the-loudest-library/{run,a80956}.py |
|---|---|
| status | executed |
| answers open question from | ana-q171-the-two-instrument-intersection-does-not-beat-a-patient-level-null-and-the-published-count-was-four-not-one |
| n libraries | 26 |
| n comparators | 18 |
| tb051939 calls | 122 |
| tb051939 rank | 12 |
| tb051939 z | 0.3 |
| median calls per library | 119 |
| max calls library | p318 |
| max calls | 189 |
| n singletons | 219 |
| tb051939 singletons | 6 |
| tb051939 singleton share | 0.0274 |
| even share | 0.0385 |
| a80956 singletons | 49 |
| a80956 singleton share | 0.2237 |
| a80956 new exclusives if dropped | 0 |
| baseline exclusives | 63 |
| median new exclusives drop one | 2 |
| worst comparator | A73051 |
| worst new exclusives | 8 |
| worst inflation PCT | 12.7 |
| pearson depth vs calls | 0.391 |
| verdict | both library-quality explanations refuted; the three single-library hits are refused as one observation, not as an artefact |
measurements — the numbers behind the chart
[
{
"quantity": "rank",
"subject": "TB051939 by gene-body calls",
"universe": "26 permutable GSE252378 libraries",
"value": 12
},
{
"quantity": "share",
"subject": "singletons owned by TB051939",
"universe": "219 genes called in exactly one library",
"value": 0.0274
},
{
"quantity": "share",
"subject": "singletons owned by A80956",
"universe": "219 genes called in exactly one library",
"value": 0.2237
},
{
"quantity": "count",
"subject": "new exclusives created by dropping A80956",
"universe": "63 baseline genome-wide exclusives",
"value": 0
},
{
"quantity": "count",
"subject": "new exclusives created by dropping the worst comparator (A73051)",
"universe": "63 baseline genome-wide exclusives",
"value": 8
},
{
"quantity": "pearson_r",
"subject": "raw depth vs gene-body call count",
"universe": "26 libraries",
"value": 0.391
}
]- What explains the ~85% of between-library variation in call count that depth does not? Something differs systematically between these libraries and it is unidentified, which is a larger uncontrolled factor than anything ruled out here.
- Leave-k-out by institution and by handling batch. Comparators may fail in correlated groups, and a set robust to every single deletion can still be fragile to the right pair.
- How many of the 63 exclusives would a 19th comparator destroy? The exclusive set may be largely a function of how many comparators happened to be sequenced, which would make 'exclusive' a property of the deposit rather than of the biology.
- ANGPT1, FOXL2 and FOXL2NB are refused as one observation rather than as an artefact. If a second deposit ever calls any of them independently, that refusal is void and they are owed entries - record the trigger now rather than rediscovering it.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article
Across the 26 permutable GSE252378 libraries the gene-body rule calls a median of 119 genes (mean 113.3, max 189 in p318). TB051939 calls 122, rank 12 of 26, z = +0.30. Of 219 genes called in exactly one library, TB051939 owns 6 (2.7%) against an even share of 3.8%, while A80956 owns 49 (22.4%). Raw depth correlates with call count at Pearson r = +0.391; TB051939 is depth rank 13 of 26.
What this evidence cannot show
Call counts are computed under this project's own occupancy rule (both replicates >=2.0x an average promoter and above the per-library max of 8 width-matched negative loci), which recovers only about a fifth of the authors' calls. A library that is quiet under this rule is not necessarily quiet under theirs, and the ranking could differ under a different threshold, which was not swept.
Evidence that argues against this (2)
- Rank 12 of 26 is a statement about the middle of a distribution whose spread is itself unexplained: depth accounts for r=+0.391, so ~85% of the variance in call count has no identified cause. 'Unremarkable among libraries that differ for unknown reasons' is a weaker exoneration than the rank makes it sound.
- The singleton share compares TB051939 to an even share of 3.8%, but singletons are not distributed evenly even in the absence of noise - a library from a distinctive patient should legitimately produce private calls. Being below even share could reflect a library that is insensitive rather than clean, and this analysis cannot separate those.
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article
Genome-wide, 63 genes are called in at least one AFF1 library and zero of the 18 comparators. Dropping each comparator in turn and recounting: the median increase is 2.0 new exclusives, the maximum is 8 (A73051, a 12.7% inflation of the baseline), and dropping A80956 - the library owning 22.4% of all singletons - yields exactly 0 new exclusives. IRX1, ANGPT1, FOXL2 and FOXL2NB each have zero comparator calls, so no single comparator deletion alters their status.
What this evidence cannot show
Leave-one-out only. Correlated failure across several comparators - by institution or handling batch - was not tested, and a set robust to every single deletion can still be fragile to a particular pair. The 63-gene baseline uses the >=1 AFF1 call threshold; at >=2 the baseline and the sensitivities both change and were not recomputed.
Evidence that argues against this (2)
- A median shift of 2 on a baseline of 63 is 3.2%, which sounds small, but the quantity that mattered in q171 was a set of 4 genes, not 63. A shift of 2 genes against a finding built on 4 is a 50% perturbation of the thing actually being argued about, and framing the sensitivity against the larger denominator understates it.
- Showing that dropping a comparator rarely CREATES exclusives says nothing about the reverse direction - how many of the 63 would be destroyed by ADDING a comparator, i.e. by a 19th patient. Since 18 comparators already remove all but 63 genes, the marginal comparator is plausibly still removing several, and the exclusive set may simply be a function of how many comparators were sequenced.
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article
An earlier version of the leave-one-out script labelled GPR18 as 'BECOMES EXCLUSIVE without A80956'. GPR18 has 0 AFF1 calls and a single comparator call from A80956, so it cannot satisfy a criterion requiring at least one AFF1 call. The label tested only whether the comparator list became empty, omitting the other half of the conjunction. The drop-one counts printed above were computed by a separate expression that tested both halves and were correct throughout, including the 0 for A80956.
What this evidence cannot show
This is a self-report of a defect in this analysis's own tooling, found and fixed in the same session. It is recorded because the label was the kind that gets quoted into a summary, and because the same conjunction error is what E1's A80956 alarm turned out to be.
Evidence that argues against this (1)
- Recording a caught display bug creates a false impression of thoroughness if the rest of the script was not audited to the same standard. It was not: only the two printed markers were re-derived by hand, and the matrix construction, normalisation and occupancy rule were reused unchanged from earlier analyses without independent re-verification in this cycle.
answers q171's second open question: TB051939 is rank 12 of 26, so the three single-library hits are one observation rather than an artefact
records the refusal q170 left owed on ANGPT1/FOXL2/FOXL2NB, on corrected grounds