Oncofusion

start / graph / ana-q180-how-far-a-baits-fixed-list-runs-into-noise-predicts-how-specific-it-looks

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q180-the-af4-specific-interactors

Across all seven MLL-fusion baits, how deep a bait's fixed top-300 list runs into non-significance predicts how many "specific" partners it appears to have — Spearman +0.821, exact p = 0.034, and MLL-AF4 is the extreme on both axes

also called depth confound in Skucha 2018 · top-300 per bait · singleton count versus max p

The published table took a fixed number of top-ranked proteins from each of seven pulldowns rather than everything above a quality bar. Baits whose fixed quota reached further down into weak, borderline hits ended up with more proteins that no other bait saw — and MLL-AF4's quota reached furthest of all. So "MLL-AF4 has the most unique partners" is mostly a statement about where its list stopped.

The same thing, for a scientist

Skucha 2018 states its own selection rule in the body: *"p-value-based filtering for the 300 most significant interactions per MLL-fusion resulted in a network of 960 high-confidence cellular proteins"* and, in the Fig. 2a legend, *"top 300 interactors per bait, ranked by p-value"* [E1]. Supplementary Table 5 carries 293 calls per bait (291 for MLL-AF4). Because n is fixed and the underlying data quality is not, the p-value AT the cut varies 10.7-fold across baits: 0.0164 for MLL-ENL up to 0.175 for MLL-AF4. Over the seven baits, the count of singleton preys (`n_fusions == 1`) tracks that cut almost perfectly: **Spearman +0.821, exact two-sided permutation p = 0.0341 over all 5,040 orderings** [E2]. Within MLL-AF4's own 291, the singletons are the weak end by every measure: median p 0.114 against 0.019 for the 42 pan-fusion proteins, 5% below p 0.05 against 83%, median 2 spectra against 9. The direction of this confound was already established in this corpus by the equal-n analysis; what is new here is the cross-bait test that quantifies it and a within-bait decomposition.

Go deeper
The seven baits, sorted by where their list stopped
baitn in Suppl. Table 5largest p in its listsingletonssingleton share
MLL-ENL2930.01647023.9%
MLL-AF92930.04786120.8%
MLL-GAS72930.06817826.6%
MLL-EEN2930.08657726.3%
MLL-AF1p2930.1077124.2%
MLL-CBP2930.11010134.5%
MLL-AF42910.17511037.8%

MLL-AF4 is last on both columns. Spearman between the two right-hand columns is +0.821; because n = 7 the permutation distribution can be enumerated exactly, and only 172 of 5,040 orderings reach |rho| >= 0.821, giving p = 0.0341 2.

This is not a claim that MLL-AF4 has no private partners. It is a claim that the count of them is not comparable across baits, because the instrument's depth is not comparable across baits, and MLL-AF4 sits at the end where the instrument is loosest.

Inside MLL-AF4: the singletons ARE the bottom of the list

Partitioning MLL-AF4's own 291 calls by how many baits saw each prey 3:

classnmedian pfraction p < 0.05median spectramean spectra
seen in 1 fusion (the "AF4-specific" set)1100.1145% (6/110)23.4
seen in 2-6 fusions1390.05546%——
seen in all 7 (the pan-fusion core)420.01983%918.0

Six of 110 reach p < 0.05, and those six are the dispatch targets this beat was sent to write. They are the top of a class whose median is p = 0.114 — i.e. the class as a whole is a non-significant tail, and the six are its least non-significant members rather than a set of hits.

A caution against over-reading even this. The p-values in the retained list are truncated at 0.175 by the rank cut, so the 5% figure is not directly comparable to a 5% type-I rate; and the GLM's p is not a monotone function of spectral count (MLLT1 under MLL-ENL is p = 9.92e-23 on 3 spectra). The comparison that carries weight is the contrast between the three rows, which uses the same statistic on the same scale in the same experiment.

Facts and figures
scriptanalyses/q180-is-af4-specificity-a-depth-artefact/run.py
statusrun
result summarySingleton count per bait tracks the largest p-value in that bait's fixed-size list at Spearman +0.821 (exact two-sided p 0.0341, 5,040 orderings). MLL-AF4 is the maximum on both. Within MLL-AF4, singletons have median p 0.114 and 2 spectra against 0.019 and 9 for the pan-fusion core.
shown for fusionYES for MLL-AF4 as one of seven baits in one experiment, in a Jurkat T-ALL host from a tagged single-copy transgene.
epigenetic readoutnone - a proteomic selection-depth analysis.
relationship to existing entriesThe direction of this finding is NOT new. ana-the-index-fusion-interactome-is-not-distinctive-and-two-earlier-answers-were-artifacts already records that the table is rank-selected per bait, that max p runs 0.0164 to 0.175, that MLL-AF4's 110/291 is a depth artefact, and that at equal n MLL-AF4 is mid-pack. This entry adds (i) the cross-bait correlation with an exact permutation p, which that entry does not have, and (ii) the within-MLL-AF4 decomposition by sharing class. It is an extension, not a rediscovery, and the equal-n result remains the stronger of the two arguments.
measurements
fraction_below_p_0_050.054fraction_below_p_0_05: 0.0545median_p0.114median_p: 0.114max_p_in_bait_list0.175max_p_in_bait_list: 0.175spearman_rho0.821spearman_rho: 0.821median_spectral_counts2.000median_spectral_counts: 2
inputs
[
  "analyses/skucha-mll-fusion-interactome/cache/41467_2018_4329_MOESM4_ESM.xlsx (Skucha 2018 Supplementary Table 5)"
]
measurements — the numbers behind the chart
[
  {
    "quantity": "spearman_rho",
    "subject": "per-bait max p | per-bait singleton count",
    "value": 0.821,
    "units": null,
    "n": 7,
    "universe": "the seven MLL-fusion baits of Skucha 2018 Supplementary Table 5",
    "null": 0,
    "note": "exact two-sided permutation p = 0.0341 over all 5,040 orderings; not sampled"
  },
  {
    "quantity": "max_p_in_bait_list",
    "subject": "MLL-AF4",
    "value": 0.175,
    "units": null,
    "n": 291,
    "universe": "Skucha 2018 Suppl Table 5, top-300-per-bait selection",
    "null": null,
    "note": "the deepest of the seven; MLL-ENL's is 0.0164, a 10.7-fold range"
  },
  {
    "quantity": "median_p",
    "subject": "MLL-AF4 singletons",
    "value": 0.114,
    "units": null,
    "n": 110,
    "universe": "MLL-AF4's own 291 calls in Skucha 2018 Suppl Table 5",
    "null": 0.019,
    "note": "null column holds the same statistic for the 42 pan-fusion proteins in the same bait"
  },
  {
    "quantity": "fraction_below_p_0_05",
    "subject": "MLL-AF4 singletons",
    "value": 0.0545,
    "units": null,
    "n": 110,
    "universe": "MLL-AF4's own 291 calls",
    "null": 0.8333,
    "note": "6 of 110; null column is 35 of 42 for the pan-fusion class in the same bait"
  },
  {
    "quantity": "median_spectral_counts",
    "subject": "MLL-AF4 singletons",
    "value": 2,
    "units": "spectra (two replicates summed)",
    "n": 110,
    "universe": "MLL-AF4's own 291 calls",
    "null": 9,
    "note": "null column is the pan-fusion class median in the same bait"
  }
]
How this could bite us back
  • The same argument applies to any future claim that KMT2A::AFF1 has a SMALLER private interactome than another fusion. Depth confounds the count in both directions, and a beat that wants to argue the index fusion is unusually promiscuous or unusually selective must equalise rank depth first.
Still open
  1. Why does MLL-AF4's list reach so much further into non-significance than the others? The obvious candidates are lower bait expression, lower purification yield, or a genuinely smaller high-confidence interactome. The bait immunoblot in Supplementary Fig. 1 would speak to the first and was not read for this entry.
  2. Does the depth-versus-singletons correlation hold in any other multi-bait AP-MS study? If it is general it is a reusable warning about top-N-per-bait designs; if it is peculiar to this table it is only a local caveat. No second dataset was checked.
Every source behind this entry (3)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    Skucha 2018 selected a fixed number of preys per bait rather than applying a p-value threshold: the body states 'p-value-based filtering for the 300 most significant interactions per MLL-fusion resulted in a network of 960 high-confidence cellular proteins' and the Fig. 2a legend states 'top 300 interactors per bait, ranked by p-value'.

    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 2018
    PMID 29777171DOI
    What this evidence cannot show

    Read from the Europe PMC full-text XML of PMC5959866 on 2026-09-24, tags stripped before entities were unescaped. The table as distributed carries 293 rows per bait, not 300, and 291 for MLL-AF4; the paper does not explain the shortfall, and the most likely reason is the documented removal of known MS contaminants (trypsin, keratin) after ranking. This entry therefore quotes the paper's '300' as the stated rule and the table's 291-293 as the delivered content, and does not reconcile them.

    Evidence that argues against this (1)
    • A fixed top-N per bait is a defensible design choice for a network figure, where an unequal number of nodes per bait would be hard to read, and the authors describe it plainly rather than hiding it. The criticism here is not of the paper's design but of a downstream use the design does not support: comparing singleton COUNTS between baits. The paper itself makes that comparison in one sentence — 'The network also revealed abundant unique interaction partners of all MLL-fusion proteins, indicating that distinct MLL-fusions can engage specific molecular pathways' — and that sentence is what this analysis contests. Nature Communications 2018
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    Over the seven MLL-fusion baits, the number of preys seen in that bait alone correlates with the largest p-value in that bait's fixed-size list at Spearman +0.821, exact two-sided permutation p = 0.0341 over all 5,040 orderings; MLL-AF4 is the maximum on both variables (0.175 and 110) and MLL-ENL / MLL-AF9 the minima.

    What this evidence cannot show

    n = 7. A Spearman on seven points has very little power and the exact p of 0.0341 comes from a permutation distribution with only 5,040 members, so it can take just a handful of discrete values near the tail. The two variables are also not independent of a third: a bait whose pulldown recovered fewer high-confidence proteins overall will both have a looser cut AND fewer chances to overlap another bait. This analysis does not separate those, and the correlation is consistent with either reading — both of which are instrumental rather than biological.

    Evidence that argues against this (1)
    • The correlation is equally consistent with the reverse causal story: a bait with genuinely many private partners would push shared preys further down its own ranking, which would raise the p at the cut. Under that reading MLL-AF4 really is the most distinctive bait and the loose cut is the CONSEQUENCE. The cross-bait correlation cannot distinguish the two. What does distinguish them is the equal-n analysis already in this corpus, which holds rank depth constant and finds MLL-AF4 mid-pack (3rd, 4th and 5th of seven at n = 50, 100 and 150) — that result, not this one, is what rules out the biological reading. This repository, entry ana-the-index-fusion-interactome-is-not-distinctive-and-two-earlier-answers-were-artifacts 2026
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    Within MLL-AF4's own 291 calls, the 110 singletons have median p 0.114 with 6 of 110 (5%) below 0.05 and median 2 spectral counts, against median p 0.019, 35 of 42 (83%) below 0.05 and median 9 spectral counts for the 42 proteins recovered under all seven baits.

    What this evidence cannot show

    The three classes are not independent samples from one population: membership in the pan-fusion class REQUIRES the prey to clear the top-300 cut in all seven baits, which selects for abundance and therefore for low p. So some of the contrast is definitional rather than informative, and the correct reading is that the singleton class is where the weak calls necessarily accumulate, not that sharing causes significance. That is still the point: a class defined so that weak calls accumulate in it should not be mined for hypotheses at face value.

    Evidence that argues against this (1)
    • The same decomposition contains a result pointing the other way, and it is in this entry's sibling: the singleton class is where DOT1L (MLL-AF9, p 0.00172, 24 spectra) and four of seven bait 3' moieties (p 1e-17 to 1e-55) live. So the class is not uniformly weak - it is bimodal, with a handful of very strong genuine positives and a long non-significant tail. MLL-AF4 happens to have none of the strong mode, because AFF1 was also seen under MLL-AF9 and so left the class. A statement that 'the singleton class is noise' would be wrong; the statement that survives is that MLL-AF4's members of it are all in the weak mode. This repository 2026
Connected to