Oncofusion

start / graph / ana-the-index-fusion-interactome-is-not-distinctive-and-two-earlier-answers-were-artifacts

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.

Three quarters of KMT2A::AFF1's top partners are shared with other MLL fusions — and the two analyses that said otherwise were both artifacts of how the source table was built

also called is the index fusion's interactome special · MLL-fusion interactome distinctiveness · rank-selected table confound

We asked whether the fusion we study grabs a distinctive set of proteins compared with six related fusions. The first two ways of measuring it gave opposite answers, and both were wrong for reasons hidden in how the published table was assembled. Measured properly, it is unremarkable: most of what it binds, its relatives bind too.

The same thing, for a scientist

**Scope of the E3 withdrawal, king-verified 2026-09-24.** The broken denominator is used ONLY in the background-check block that produces E3. **E1 (rank-based selection) and E2 (equal-n distinctiveness) are computed from `ranked` / `maxp` / `unique_counts` and never touch `neg_ctrl` or the summed `signal` variable.** The title's claim — that the two earlier analyses were artifacts of how the source table was built — rests on E1, which stands. **The thesis of this entry is unaffected by the withdrawal.** Skucha 2018's Supplementary Table 5 (PMID 29777171) gives seven MLL-fusion AP-MS baits in one isogenic Jurkat system, so it permits a within-experiment comparison of how distinctive each fusion's interactome is. The question bears on the menin-interface hypothesis: if MLL-AF4's partners are overwhelmingly shared, the retained KMT2A N-terminus dominates the interactome and the AFF1 C-terminus contributes comparatively little. **The measurement is confounded twice over, in opposite directions.** As listed, MLL-AF4 looks the MOST distinctive of the seven (110/291, 37.8% unique) — an artifact, because the table is **rank-selected per bait** (each carries ~293 calls while the largest p ranges 0.0164 for MLL-ENL to 0.175 for MLL-AF4), so MLL-AF4's column reaches deepest into the weak tail. Imposing a common p threshold **overcorrects** and makes it the LEAST distinctive (3/50, 6.0% at p<0.0164), because it leaves MLL-AF4 with 50 partners against MLL-ENL's 292 and a smaller set drawn from a shared core has fewer chances to contain anything private. **Only equalising n removes both.** At the top 50, 100 and 150 by rank per bait, MLL-AF4 is unremarkable — 26.0%, 21.0%, 23.3% unique, ranking 3rd, 4th and 5th of seven. MLL-AF9 is consistently least distinctive and MLL-CBP most (2 of 3 sizes).

Go deeper
Why this is filed even though the answer is negative
What this does to the menin-interface hypothesis
Facts and figures
scriptanalyses/mll-fusion-interactome-distinctiveness/run.py
resultsanalyses/mll-fusion-interactome-distinctiveness/results.json
shown for fusionYES, MLL-AF4 specifically, as one of seven baits in a single experiment — but in a Jurkat T-ALL host from a tagged single-copy transgene, not the endogenous fusion in B-lineage disease.
three answers
common threshold p < 0.01…6.000common threshold p < 0.0164: 6top 100 by rank per bait21.0top 100 by rank per bait: 21top 150 by rank per bait23.3top 150 by rank per bait: 23.3top 50 by rank per bait26.0top 50 by rank per bait: 26as listed in the table37.8as listed in the table: 37.8
source
{
  "pmid": "29777171",
  "table": "Supplementary Table 5 (MOESM4), 960 proteins, seven baits, one isogenic Jurkat system",
  "proteomexchange": "PXD009338"
}
the confound
{
  "evidence_that_selection_is_rank_based": "per-bait call counts are 293, 291, 293, 293, 293, 293, 293 while the largest p-value per bait ranges from 0.0164 (MLL-ENL) to 0.175 (MLL-AF4). Near-identical counts with an order-of-magnitude spread in effective cutoff is a fixed top-N per bait, not a shared significance threshold.",
  "direction_of_bias_uncorrected": "inflates MLL-AF4's apparent uniqueness, because its column reaches furthest into the weak tail where partners are least likely to be independently recovered in another bait",
  "direction_of_bias_at_common_threshold": "deflates it, because equalising p leaves MLL-AF4 with 50 partners against MLL-ENL's 292, and uniqueness fraction depends on set size when a shared core exists",
  "stated_by_the_authors_verbatim": "Methods: \"All the identified bait-prey pairs were ranked by the combined p-value. The 300 most significant interactions per bait were retained.\" Fig. 2 legend: \"960 candidate genes (top 300 interactors per bait, ranked by p-value)\". Found by Q84; verified by me from PMC5959866."
}
three answers — the numbers behind the chart
[
  {
    "method": "as listed in the table",
    "mll_af4_unique_pct": 37.8,
    "rank_of_7": "1st (most distinctive)",
    "verdict": "ARTIFACT"
  },
  {
    "method": "common threshold p < 0.0164",
    "mll_af4_unique_pct": 6,
    "rank_of_7": "7th (least distinctive)",
    "verdict": "ARTIFACT"
  },
  {
    "method": "top 50 by rank per bait",
    "mll_af4_unique_pct": 26,
    "rank_of_7": "3rd",
    "verdict": "sound"
  },
  {
    "method": "top 100 by rank per bait",
    "mll_af4_unique_pct": 21,
    "rank_of_7": "4th",
    "verdict": "sound"
  },
  {
    "method": "top 150 by rank per bait",
    "mll_af4_unique_pct": 23.3,
    "rank_of_7": "5th",
    "verdict": "sound"
  }
]
stable across sizes
{
  "least_distinctive": "MLL-AF9, at all three sizes (10, 10, 25 unique)",
  "most_distinctive": "MLL-CBP at 50 and 100; MLL-GAS7 at 150",
  "index_fusion": "ranks 3rd, 4th, 5th — no stable position, which is itself the result"
}
permutation
{
  "design": "2,000 draws of 100 proteins per bait, uniformly from the observed 960-protein union, all seven baits randomised",
  "observed_mll_af4_unique": 21,
  "null_mean": 51.8,
  "null_sd": 4.8,
  "one_sided_p": 0.0005,
  "what_it_does_and_does_not_show": "It shows the seven baits share far more than independent uniform sampling would produce, i.e. there is a real common core — which the paper already reported as 42 pan-fusion proteins. It does NOT test whether any bait differs from any other, because the null has no core to begin with. It is a floor, not a comparison."
}
background check
{
  "pan_fusion_with_zero_control": "32 of 42",
  "nonzero_ones": "ACTB 6, PSIP1 5, CHD4 4, SF3A3 3, MCM2 2, CSNK2A1 2, RUNX1 2, ZNF512 2, TCEB3 1, SMARCC1 1",
  "note": "PSIP1 is LEDGF, an established KMT2A partner, and carries control signal 5 - a reminder that non-zero control does not mean false positive, only unresolved.",
  "control_per_unit_signal_WITHDRAWN": {
    "pan_fusion": 0.0023,
    "singletons": 0.0172,
    "permutation_two_sided_p": 0.0694,
    "withdrawn": "2026-09-24 Q180: denominator summed over all 7 baits against a 1-column numerator"
  },
  "control_per_unit_signal_CORRECTED": {
    "pan_fusion": 0.01625,
    "singletons": 0.01723,
    "fold": 1.06,
    "permutation_two_sided_p": 0.9102,
    "nperm": 20000,
    "seed": 180,
    "mean_denominator_pan_fusion": 45.1,
    "mean_denominator_singletons": 15.7,
    "note": "no discrimination; the negative-control column cannot separate core from tail"
  },
  "raw_detectability_runs_the_other_way": {
    "pan_fusion_nonzero_control": "10/42 (23.8%)",
    "singletons_nonzero_control": "68/568 (12.0%)"
  },
  "e1_e2_verified_unaffected": "king-verified 2026-09-24: the summed-signal denominator appears only in the background-check block; the rank and equal-n analyses are independent of it"
}
How this could bite us back
  • This entry supports no clinical action and names no target. It is a comparison of published partner lists and an audit of two of my own analyses.
  • Nothing here should be read as evidence that any protein on any of these lists is a real KMT2A::AFF1 interactor in infant B-lineage leukaemia. The source is a tagged transgene in a T-ALL cell line, and this project has previously mistaken heterologous-system preys for established partners.
Still open
  1. RE-OPENED 2026-09-24 (Q180). This was marked RESOLVED on the strength of a 7.5-fold cleaner-per-unit-signal figure that is now WITHDRAWN as a denominator artefact (corrected fold 1.06, p 0.9102). What still stands: 32 of 42 pan-fusion proteins have zero control signal, so there is no positive evidence the core IS background. What does not: the claim that the shared fraction is no longer only an upper bound on shared biology. On the negative-control axis the core and the tail are indistinguishable, so that axis cannot license the upgrade.
  2. Does MLL-CBP's higher distinctiveness reflect real biology? CBP is a histone acetyltransferase that brings its own enzymatic machinery, so a larger private interactome is the predicted result — and if so it is a positive control that this measurement detects real partner-driven differences rather than noise.
  3. The comparison that would actually settle the N-terminus question is not available here: all seven baits SHARE the KMT2A N-terminus, so a shared core is the expected result under either hypothesis. The discriminating experiment is a bait carrying the AFF1 C-terminus WITHOUT the KMT2A N-terminus, or full-length AFF1, against the fusion in the same system.
  4. Repeat the equal-n analysis on the 259 MLL-AF4 partners this graph does not hold, to see whether the missing ones are preferentially in the shared core or the private tail. If they are mostly private and weak, the coverage gap matters less than the raw 11% suggests.
Every source behind this entry (3)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    In Skucha 2018's seven-bait AP-MS table the per-bait call counts are near-identical (293 for six baits, 291 for MLL-AF4) while the largest p-value per bait ranges from 0.0164 for MLL-ENL to 0.175 for MLL-AF4. 186 of MLL-AF4's 291 listed partners have p > 0.05. The selection is therefore a fixed number of top-ranked preys per bait, not a common significance threshold, and any coverage or uniqueness figure taken off the raw listing inherits that.

    MLL-fusion-driven leukemia requires SETD2 to safeguard genomic integrity. · Skucha A, Ebner J, Schmoellerl J, Roth M, Eder T, Cesar-Razquin A, Stukalov A, Vittori S, Muhar M, Lu B, Aichinger M, Jude J, Mueller AC · Nature Communications 2018
    PMID 29777171DOI
    What this evidence cannot show

    CORRECTED 2026-09-21 (cycle 47). This caveat previously read 'This is my inference from the distribution of the table's own values, not a statement the paper makes about its selection procedure; the Methods describe the generalised linear model but I did not find text specifying a fixed per-bait cutoff.' THAT WAS FALSE, and the failure was mine: I grepped the full text for the terms I expected rather than reading the filtering section. The Methods state it in one sentence - "All the identified bait-prey pairs were ranked by the combined p-value. The 300 most significant interactions per bait were retained. This cutoff represents a compromise between ensuring high statistical significance of the included interactors, but also capturing s[ensitivity]" - and the Fig. 2 legend repeats it as "top 300 interactors per bait, ranked by p-value". So the rank selection is the authors' stated, deliberate and reasoned design, not a property I deduced. The inference was correct; presenting it as an inference was not. Verified against PMC5959866 by me on 2026-09-21. The recomputation of the paper's own 128 and 42 still passes exactly.

    Evidence that argues against this (1)
    • Reproducing the paper's own summary counts validates the parsing, not the stringency — the 128 and 42 are computed from the same listing and inherit the same rank selection. So the controls that make this extraction trustworthy say nothing about whether the listed partners are individually reliable, and it would be a mistake to treat passing controls as evidence the weak tail is real. Nature Communications 2018
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    Equalising the number of partners per bait, MLL-AF4 is not distinctive among the seven MLL fusions in either direction. At the top 50, 100 and 150 by rank it has 26.0%, 21.0% and 23.3% partners unique to it, ranking 3rd, 4th and 5th of seven. MLL-AF9 is the least distinctive at all three sizes and MLL-CBP the most at two of three. Roughly three quarters to four fifths of the index fusion's top-ranked partners are shared with at least one other MLL fusion.

    What this evidence cannot show

    Rank order across baits is NOT stable for the index fusion (3rd, 4th, 5th across three sizes), so the correct reading is that the baits are largely indistinguishable at equal n rather than that MLL-AF4 sits at a particular position. All seven baits come from one laboratory, one cell line and one instrument, so shared partners may partly reflect shared background rather than shared biology — the GLM's DDX5/DDX17 normalisation mitigates this but cannot remove it. Jurkat is a T-ALL host and the baits are tagged single-copy transgenes.

    Evidence that argues against this (1)
    • A uniqueness fraction computed over one experiment's baits cannot distinguish shared BIOLOGY from shared ARTEFACT: proteins that stick to a Strep-HA tag, to the nuclear-lysate preparation, or to the beads would appear in all seven columns and register as a pan-fusion core. The paper's negative controls address this in principle, and the neg_ctrl column exists in the table, but this analysis did not use it. Until the pan-fusion core is checked against that column, the shared fraction is an upper bound on shared biology. Nature Communications 2018
  3. E3weak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.database

    The shared pan-fusion core is not explained by bead or tag background. Using the table's own neg_ctrl column, 32 of the 42 pan-fusion proteins have zero control signal. The 10 that do not are at low magnitude (1 to 6 counts) and include ACTB, the classic sticky protein, at 6. Pan-fusion proteins do show non-zero control signal MORE often than singletons (23.8% vs 12.0%), but they also carry 20-fold more spectral counts (mean 315.6 vs 15.7), and detectability in a control is a property of abundance. 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). **[WITHDRAWN 2026-09-24, Q180 — the normalisation was broken.] The '7.5-fold LOWER' figure is an artefact of the denominator.** `analyses/mll-fusion-interactome-distinctiveness/run.py` line 156 computes `signal = sum(_counts(r[c+1]) for c in BAIT_COL.values())`, summing spectral counts over ALL SEVEN baits, while the numerator is the single `neg_ctrl` column. A pan-fusion prey is non-zero in seven bait columns and a singleton in one, so the denominator is structurally ~7x larger for the core BY CONSTRUCTION. The '20-fold more spectral counts' quoted above is itself the artefact: mean denominators 315.6 vs 15.7 = 20.16 = **7 x 2.88 exactly**, where 2.88 is the genuine abundance difference and 7 is the summation. Recomputed with a per-bait denominator: **pan-fusion 0.01625 vs singletons 0.01723, fold 1.06, two-sided permutation p = 0.9102** (20,000 draws, seed 180). **There is no difference.** King-verified: the reproducer reproduces the published 0.00232 / 0.01723 / 7.42 / p 0.0745 exactly before correcting them. Reproducer: `analyses/q180-is-af4-specificity-a-depth-artefact/control_denominator.py`. **What survives:** the RAW counts, which are unaffected — 32 of 42 pan-fusion proteins have zero control signal, and the non-zero ten are at 1-6 counts. **What does not:** any claim that the core is CLEANER per unit signal. And the raw direction runs the other way — pan-fusion proteins carry non-zero control signal MORE often than singletons (23.8% vs 12.0%), which this item already reported and then normalised away with a broken denominator.

    What this evidence cannot show

    The permutation gives two-sided p = 0.0694, which does NOT reach significance at 0.05. The correct statement is that the core is not dirtier than the tail, and directionally cleaner; it is not established that it is cleaner. 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. The neg_ctrl column's own composition — how many control pulldowns, and of what — was not read from the Methods.

    Evidence that argues against this (2)
    • The normalisation in this item is withdrawn: its denominator summed spectral counts over all seven baits against a single-column numerator, giving the pan-fusion core a 7x larger denominator by construction. Corrected, the fold is 1.06 and p = 0.9102, so the negative-control column does not discriminate the core from the tail at all. The raw zero-control count (32 of 42) is unaffected and still supports the item's headline.
    • Absence of control signal cannot distinguish a genuine interactor from a protein that co-purifies for a structural reason unrelated to the bait, such as a component of a complex that binds the tag's neighbourhood or a protein retained by the nuclear-lysate preparation itself. A clean negative control rules out bead and free-tag stickiness, which is the narrowest of the contamination classes. ACTB at 6 counts remains in the core and should be treated as unresolved rather than as a partner. Nature Communications 2018
Connected to