This repo's "the pan-fusion core is 7.4-fold cleaner per unit signal than the singleton tail" divides one control column by a denominator summed over seven baits — on a per-bait denominator the ratio is 1.06 and p goes from 0.074 to 0.91
also called neg_ctrl per unit signal · denominator artefact · mll-fusion-interactome-distinctiveness background_check
An earlier analysis here concluded that the proteins shared by all seven fusions are far cleaner than the fusion-specific ones, by dividing how much each protein appears in the negative control by how much it appears overall. But "overall" added up all seven pulldowns for a shared protein and only one for a specific protein, so the shared proteins were handed a denominator seven times too big. Put on the same footing, the two groups are indistinguishable.
The same thing, for a scientist
`analyses/mll-fusion-interactome-distinctiveness/run.py` line 156 computes each prey's signal as `sum(_counts(r[c+1]) for c in BAIT_COL.values())` — spectral counts summed over all seven bait columns — and divides the single `neg_ctrl` column by it. A prey present in seven baits therefore gets a denominator built from seven non-empty columns; a prey present in one gets six zeros and one column. **The reported 7.4-fold difference (pan-fusion 0.00232 vs singleton 0.01723) is almost exactly the seven-fold structural asymmetry**, and the decomposition is exact: the mean denominators are 315.6 and 15.7, a 20.16-fold ratio, which factorises as 7 baits x 2.88-fold genuine per-bait abundance. Recomputing with the denominator divided by the number of baits the prey was called in — the same numerator, the same permutation, seed 180, 20,000 draws — gives **pan-fusion 0.01625 vs singleton 0.01723, fold 1.06, two-sided p = 0.9102** [E1]. The published figures replicate exactly under the published method, so this is a specification defect, not an execution error.
Both computations, side by side
| denominator | pan-fusion (n=42) | singletons (n=568) | fold | permutation p | |
|---|---|---|---|---|---|
| as published | spectral counts summed over all 7 baits | 0.00232 | 0.01723 | 7.42 | 0.0745 |
| control | the same, divided by the number of baits the prey was called in | 0.01625 | 0.01723 | 1.06 | 0.9102 |
Mean denominators: as published 315.6 vs 15.7 (20.16x); per bait 45.1 vs 15.7 (2.88x). 20.16 = 7 x 2.88 to two decimals — the structural factor and the genuine abundance factor separate cleanly 1.
The published point estimates 0.00232 and 0.01723 reproduce to five decimal places under the published method, and the sheet's own n_fusions column agrees with the count of non-empty bait p-value columns for all 960 rows, so there is no data disagreement between the two analyses.
What survives and what does not
Survives. The headline verdict of the earlier analysis — "No evidence the shared core is bead or tag background" — is unchanged, and if anything is on firmer ground: under the corrected denominator the two classes are indistinguishable (p = 0.91) rather than differing at a non-significant p = 0.069. The raw counts it also reported are unaffected: 32 of 42 pan-fusion proteins have zero control signal, and the non-zero ones are at 1-6 counts.
Does not survive. Three specific statements:
- "the core is 7.5-fold cleaner per unit signal than the singleton tail" — the fold is 1.06.
- "Normalised for signal the core is cleaner, not dirtier" — it is neither.
- The
open_questionsline inana-the-index-fusion-interactome-is-not-distinctive-and-two-earlier-answers-were-artifactsreading "RESOLVED this cycle, see E3 ... The shared fraction is therefore no longer only an upper bound on shared biology" — the resolution rested on the 7.5-fold figure. With the normalised comparison flat, the neg_ctrl column cannot discriminate the two classes at all, and the upper-bound caveat that line retired should be reinstated.
That entry's own caveat came one step short of this: it says "The ratio metric also divides by a quantity that partly determines the numerator, so it is a heuristic for stickiness rather than a principled model." True, and not the defect — the defect is that the denominator's DEFINITION differs between the two arms being compared.
For this beat's purposes the consequence is the one that matters: the neg_ctrl column gives no evidence that the AF4-specific tail is dirtier than the shared core. The case against that tail has to be made on compartment, on spectral count and on selection depth — as the sibling entries do — and not on the negative control.
Reported, not edited
Per the standing rule, this is a report on a file this beat does not own. The affected artefacts are analyses/mll-fusion-interactome-distinctiveness/run.py (line 156 and the background_check block of its results.json) and evidence item E3 plus the first open_questions entry of ana-the-index-fusion-interactome-is-not-distinctive-and-two-earlier-answers-were-artifacts.
The sentence that should change, in E3: "Normalised as control signal per unit of observed signal, the core is 7.5-fold LOWER than the singleton tail (0.0023 vs 0.0172)" should read "Normalised per bait in which the prey was called, the two classes are indistinguishable (0.0163 vs 0.0172, permutation p = 0.91); the 7.5-fold figure previously reported here divided by a denominator summed over seven baits and is a structural artefact." The reproducing script is analyses/q180-is-af4-specificity-a-depth-artefact/control_denominator.py.
| script | analyses/q180-is-af4-specificity-a-depth-artefact/control_denominator.py |
|---|---|
| status | run |
| result summary | Published metric reproduced exactly (0.00232 / 0.01723, fold 7.42, p 0.0745). Per-bait denominator gives 0.01625 / 0.01723, fold 1.06, p 0.9102. The 20.16-fold denominator gap factorises as 7 x 2.88. |
| shown for fusion | NO - nobody has looked in any fusion. This is a statistics-of-the-source-table finding and involves no fusion biology. |
| epigenetic readout | none - a normalisation audit of a proteomic contamination metric. |
inputs
[ "analyses/skucha-mll-fusion-interactome/cache/41467_2018_4329_MOESM4_ESM.xlsx", "analyses/mll-fusion-interactome-distinctiveness/run.py (the metric being audited)", "analyses/mll-fusion-interactome-distinctiveness/results.json (the figures being reproduced)" ]
measurements — the numbers behind the chart
[
{
"quantity": "ratio_neg_ctrl_per_signal",
"subject": "pan-fusion preys, denominator summed over 7 baits",
"value": 0.00232,
"units": null,
"n": 42,
"universe": "all 960 rows of Skucha 2018 Suppl Table 5, classes n_fusions==7 and n_fusions==1",
"null": 0.01723,
"note": "null column is the singleton class under the SAME published metric; fold 7.42, permutation p 0.0745. This is the figure being corrected, reproduced exactly."
},
{
"quantity": "ratio_neg_ctrl_per_signal",
"subject": "pan-fusion preys, denominator per bait called in",
"value": 0.01625,
"units": null,
"n": 42,
"universe": "same 960 rows, same classes, same numerator, denominator divided by n_fusions",
"null": 0.01723,
"note": "null column is the singleton class under the corrected metric; fold 1.06, two-sided permutation p 0.9102 over 20,000 draws, seed 180"
},
{
"quantity": "fold_change",
"subject": "mean denominator, pan-fusion over singleton, as published",
"value": 20.16,
"units": null,
"n": 610,
"universe": "the 42 pan-fusion and 568 singleton rows of Skucha 2018 Suppl Table 5",
"null": 7,
"note": "null column is the purely structural factor (seven bait columns vs one); the residual 2.88 is the genuine per-bait abundance difference, and 7 x 2.88 = 20.16"
}
]corrects
{
"analysis_dir": "analyses/mll-fusion-interactome-distinctiveness",
"code_line": "run.py line 156: signal = sum(_counts(r[c + 1]) for c in BAIT_COL.values())",
"entry": "ana-the-index-fusion-interactome-is-not-distinctive-and-two-earlier-answers-were-artifacts",
"fields": [
"evidence E3 claim",
"open_questions[0] ('RESOLVED this cycle, see E3')",
"data.background_check.control_per_unit_signal"
],
"action": "reported to the coordinator, not edited - this beat does not own those files"
}- Nothing here touches an intervention directly, but a contamination metric that reads clean when it is structurally biased is the kind of thing that licenses a partner into a target list. The 7.5-fold figure had already been used once to retire an upper-bound caveat on how much of the pan-fusion core is shared biology.
- How many control purifications does the neg_ctrl column summarise, and of what cells? The PRIDE record for PXD009338 lists the raw files; counting the control runs there would turn both analyses' absolute ratios from uninterpretable into interpretable, and is a listing call away.
- Are there other metrics in this repository that divide a fixed numerator by a denominator whose DEFINITION differs between the two arms being compared? This one survived a cycle with a caveat attached that named a different, lesser problem with the same metric.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
Replicating this repository's neg_ctrl-per-signal metric exactly reproduces the published pan-fusion 0.00232 and singleton 0.01723 (fold 7.42, permutation p 0.0745); substituting a per-bait denominator gives 0.01625 and 0.01723 (fold 1.06, two-sided permutation p 0.9102 over 20,000 draws, seed 180), and the mean-denominator ratio of 20.16 factorises exactly as 7 baits times a 2.88-fold genuine per-bait abundance difference.
Denominator control for analyses/mll-fusion-interactome-distinctiveness, run 2026-09-24 by Q180. Script analyses/q180-is-af4-specificity-a-depth-artefact/control_denominator.py, output control_denominator.json. Asserts the bait column map and the neg_ctrl column against the sheet's header rows, and asserts that the sheet's n_fusions column equals the count of non-empty bait p-value columns for all 960 rows before computing anything. · This repository (script and JSON committed; deterministic, seeded, no network) 2026What this evidence cannot show
The per-bait denominator is a choice, not the uniquely correct normalisation. It puts the two arms on the same per-experiment footing, which is what the comparison needs, but a prey recovered in seven baits genuinely was observed seven times and a metric that ignores that discards information. A fully principled version would model the control as a Poisson rate against per-bait exposure rather than taking a ratio of sums at all; neither analysis does that. What is not a matter of choice is that the ORIGINAL denominator differs in definition between the two arms being compared, which no normalisation choice can justify.
Evidence that argues against this (1)
- The correction does not rescue the opposite claim either. Under the corrected metric the singleton tail is not shown to be DIRTIER than the core - p = 0.9102 is a flat result, and flat means the neg_ctrl column simply cannot separate these classes at this n. A reader wanting to argue that the AF4-specific tail is contamination gains nothing from this entry and should rely on the compartment audit instead. It is also worth recording that the earlier analysis's directional conclusion ('no evidence the shared core is bead or tag background') is UNCHANGED by the correction; only its magnitude and its stated open-question resolution fail. This repository, entry ana-the-index-fusion-interactome-is-not-distinctive-and-two-earlier-answers-were-artifacts, evidence item E3 2026
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
Skucha 2018's Supplementary Table 5 carries exactly one neg_ctrl column of two replicate spectral counts for the whole table, shared across all seven baits, so any per-bait normalisation of it is an assumption rather than a measurement.
MLL-fusion-driven leukemia requires SETD2 to safeguard genomic integrity. · Skucha A, Ebner J, Schmöllerl J, Roth M, et al. (20 authors; Grebien F senior author) · Nature Communications 2018What this evidence cannot show
Neither the paper's Methods nor the PRIDE record for PXD009338 states how many control purifications the neg_ctrl column summarises, or from what - mock-transfected Jurkat, untagged parental, or Dox-untreated transgenic cells. The paper's model estimates a baseline abundance term 'from the control AP-MS experiments', plural, but does not enumerate them. Both this analysis and the one it corrects treat the column as a single comparable quantity because nothing else is available, and that assumption is unverified.
Evidence that argues against this (1)
- If the neg_ctrl column is itself a SUM over several control runs, the numerator has an exposure problem of its own that neither analysis addresses, and the per-bait denominator does not fix it. In that case the absolute level of both ratios is uninterpretable and only their comparison between classes carries information - which is precisely the comparison this entry finds flat. So the correction's conclusion is robust to this unknown while both analyses' absolute numbers are not. ProteomeXchange / PRIDE, EMBL-EBI 2018
its E3 figure of '7.5-fold cleaner per unit signal' and the open-question resolution built on it; the headline verdict that the core is not bead or tag background is unaffected
having removed the negative-control column as a discriminator, the compartment audit is the only quality axis left that separates the two prey classes
the compartment audit carries the quality argument because the negative-control column was shown there to be unable to separate the two prey classes