Oncofusion

start / graph / ana-proximity-weak-spot-readjudication

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q20-proximity-interactome

Re-adjudicating the standing weak spot: proximity labelling HAS sampled AFF1, KMT2A and five leukaemia fusions - just never a KMT2A fusion

also called weak spot 1 readjudication · transient interactome re-audit

The project has carried a note saying nobody has ever used modern neighbour-tagging methods on this biology. Half of that is wrong: it has been done on AFF1, on KMT2A and on five other leukaemia fusion proteins. The half that is right is that nobody has done it on a KMT2A fusion.

The same thing, for a scientist

The standing weak spot reads: *'The transient interactome has never been sampled. Two independent queries for BioID/TurboID of any KMT2A fusion return zero. Every composition claim rests on 2004-2012 affinity purification. A single modern proximity map could overturn several negatives at once, including the WRAD one.'* **Sentence two is confirmed and is the only sentence that survives intact.** Sentence one is **wrong**: the curated proximity record holds 4 rows for AFF1, 5 for AFF4, 12 for ELL, 9 for MLLT1, 14 for CDK9, 16 for CCNT1, 46 for KMT2A, 6 for DOT1L and 4 for MEN1, across five publications [E1]. Sentence three is **narrowed**: composition claims about the fusion still rest on affinity purification, but wild-type neighbourhood data exist. Sentence four is **over-optimistic** and should be struck: in the one dataset available, obligate complexes do not recover as blocks, so a proximity map would add a list rather than adjudicate the WRAD negative [E2][E4].

Go deeper
The verdict, sentence by sentence
sentenceverdict
'The transient interactome has never been sampled.'Wrong. Five publications contribute proximity records for these proteins; AFF1 and KMT2A are both in them 1.
'Two independent queries for BioID/TurboID of any KMT2A fusion return zero.'Confirmed, and reproduced with thirteen queries 3. No proximity experiment has ever used a KMT2A fusion, or any KMT2A-complex protein, as a bait 1.
'Every composition claim rests on 2004-2012 affinity purification.'True of the fusion; not true of the wild-type proteins.
'A single modern proximity map could overturn several negatives at once, including the WRAD one.'Over-optimistic. Proximity and affinity purification recover few shared partners by design 2, and in the available dataset WRAD splits WDR5 11 / ASH2L 9 / RBBP5 0 / DPY30 0 across the same fifty baits 4. A proximity map would produce a different list, not a verdict on the old one.
Why the original audit missed it, and the rule that follows

The weak spot was written from gene-symbol-plus-method-name PubMed queries, and those queries are genuinely empty. BioID AND MLL returns 0. BioID AND KMT2A returns 0. (BioID OR TurboID OR APEX2 OR "proximity labeling" OR "proximity labelling" OR "proximity-dependent biotinylation" OR "proximity biotinylation" OR BioID2 OR "split-TurboID") AND (KMT2A OR MLL OR MLL1 OR "mixed lineage leukemia" OR "mixed-lineage leukemia") returns 0, with the query translations checked so the zeros are not a parser artefact 3.

The data are nonetheless there, because large-scale interactome papers do not name their preys. The strings AFF1, AFF4, ELL and KMT2A appear nowhere in the body text of the paper that contains AFF1's entire modern proximity record; they exist only in its supplementary files and its IMEx deposition 5.

This is the project's standing rule 19 - a symbol-driven absence cannot see a complex-level abstract - in a sharper form. A symbol-driven absence cannot see a prey. The generalisation: for any protein, the question 'has anyone measured X about it' cannot be answered from titles and abstracts when X is measured at scale by studies about something else. The entry point has to be the method code in a structured database, here detmethod:"MI:1314" in IntAct.

What is genuinely still missing
  1. No KMT2A fusion has ever been proximity-labelled. Confirmed 13.
  2. No KMT2A-complex protein has ever been a bait - not KMT2A, not menin, not AFF1, AFF4, ELL, ELL2, MLLT1, MLLT3, MLLT10, DOT1L or CDK9. Every fact in this beat is one-directional 1.
  3. AF9/MLLT3 and ELL2 appear in no proximity record at all, from any direction 1.
  4. Nothing has been done in a KMT2A-rearranged cell. The AFF1 data are K562, which is BCR::ABL1 5.

And one thing that is not missing: the protocol. TurboID has been fused in frame to three AML oncofusion cDNAs and run in primary murine haematopoietic cells, and BioID to two nucleoporin fusions. The gap is selective, not technical.

Facts and figures
scriptnone (database and literature queries)
statusrun
n pubmed queries13
n intact queries14
result summaryHalf the weak spot is wrong, half is confirmed, and the reason it was wrong is that a symbol-driven PubMed query cannot see a prey in a large-scale dataset. The entry point must be the PSI-MI method code in a structured database.
verdict
{
  "sentence_1_transient_interactome_never_sampled": "WRONG",
  "sentence_2_bioid_turboid_of_any_kmt2a_fusion_returns_zero": "CONFIRMED",
  "sentence_3_composition_claims_rest_on_2004_2012_ap": "TRUE OF THE FUSION ONLY",
  "sentence_4_one_map_could_overturn_several_negatives_including_wrad": "OVER-OPTIMISTIC - STRIKE"
}
still missing
[
  "no KMT2A fusion proximity-labelled",
  "no KMT2A-complex protein ever a bait",
  "MLLT3 and ELL2 absent from the whole proximity record",
  "nothing done in a KMT2A-rearranged cell"
]
How this could bite us back
  • Correcting a weak spot makes it look solved. It is not: the fusion-side gap is real, total, and is the thing the weak spot was pointing at.
  • The existence of wild-type AFF1 proximity data creates a new hazard - someone will cite it for the fusion. Every entry in this folder says it cannot be, which is exactly the kind of caveat that does not survive being summarised.
Still open
  1. How many other standing weak spots in this project were written from symbol-plus-method PubMed queries and are wrong the same way? The audit is cheap: re-run each against a structured database keyed on method rather than topic.
  2. Should the coverage ledger's weak spot 1 be rewritten rather than deleted? The confirmed half is more valuable stated precisely than the original was stated broadly.
  3. Does BioGRID, which curates proximity data under its own method vocabulary, hold rows IntAct does not? Its REST API requires an access key that was not available in this session.
Every source behind this entry (5)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    The curated IntAct proximity record (detection method MI:1314) holds 46 rows for KMT2A, 16 for CCNT1, 14 for CDK9, 12 for ELL, 9 for MLLT1, 6 for DOT1L, 5 for AFF4, 4 for AFF1, 4 for MEN1 and 4 for MLLT10, drawn from five publications, and none at all for MLLT3 or ELL2; in every row the protein is in the prey role, and no super elongation complex component, KMT2A-complex component or KMT2A fusion has ever been used as a proximity-labelling bait. The method filter passes its control with 64,945 MI:1314 rows in total and 259 for LMNA.

    What this evidence cannot show

    IntAct is one database with one curation policy. The Human Cell Map (192 BioID baits in HEK293) is not served through this route and was not queried, because no REST endpoint was found on 2026-09-19. The claim is therefore bounded to the curated IntAct record.

    Evidence that argues against this (1)
    • A database-bounded census can only refute an absence claim, never confirm one. The bait-side zero would be overturned by a single unpublished or uncurated experiment, and this beat found no way to rule that out. Go CD et al. 2021
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    Proximity biotinylation and affinity purification of chromatin-associated complexes identified few shared interaction partners and enriched for distinct biological processes, with both recovering biologically meaningful interactions.

    PMID 25281560DOI
    What this evidence cannot show

    HEK293, histones and Mediator. It is the basis for striking the weak spot's fourth sentence, and it is a methods comparison rather than a result about KMT2A.

    Evidence that argues against this (1)
    • The paper's conclusion is complementarity. Low overlap means a proximity map would add information, so 'it would not overturn the WRAD negative' is not the same as 'it would not be worth doing'. Lambert JP et al. 2015
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    Thirteen PubMed queries combining proximity-labelling vocabulary with KMT2A, MLL, MLL1, mixed lineage leukemia, AFF1, AF4, MLLT2, AFF4, ELL, ELL2, EAF1, EAF2, MLLT1, ENL, MLLT3, AF9, MLLT10 and AF10 return zero or near-zero records; BioID AND MLL returns 0, BioID AND KMT2A returns 0, and the pooled nine-term method OR five-term disease query returns 0, while the control terms return BioID 853, proximity labeling 1,455, KMT2A 3,216, MLL 7,520, BioID AND leukemia 14 and TurboID AND leukemia 12. Query translations were read to confirm the zeros are not a parser artefact.

    What this evidence cannot show

    Run 2026-09-19. A counted absence is only as good as its symbol set; the aliases used are listed in the enumeration entry. The SEC term is a contaminated alias (it matches Sec bodies, Sec61 and seconds) and its 12 records were title-checked to zero on-beat. These zeros bound the TITLE AND ABSTRACT literature and, as this beat demonstrates, cannot see a prey in a large-scale dataset.

    Evidence that argues against this (1)
    • The same searches that return zero coexist with 46 curated KMT2A proximity rows in IntAct, so this counted absence is a demonstration that the query class is blind rather than evidence that the data do not exist. It is recorded as the negative control for the beat's method, not as a finding about the literature. IntAct molecular interaction database (EMBL-EBI) 2026
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    In the K562 50-bait BioID2 map, nominally obligate complexes do not recover as blocks: WDR5 11 baits, ASH2L 9, RBBP5 0, DPY30 0; CDC73 25, RTF1 15, PAF1 6, CTR9 3, LEO1 0.

    What this evidence cannot show

    Recovery is abundance- and size-weighted and every zero is a zero after the authors' bona fide filter, so the split is a detectability statement. It nonetheless bears directly on whether a proximity map could adjudicate the WRAD negative, which is what the weak spot claimed.

    Evidence that argues against this (1)
    • In the largest human BioID survey a prey's bait profile is dominated by subcellular compartment rather than binding, so any neighbourhood comparison must control for localisation before it means anything. Go CD et al. 2021
  5. E5established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    The paper containing AFF1's entire modern proximity record does not mention AFF1, AFF4, ELL or KMT2A anywhere in its body text; those proteins are reachable only through its supplementary data files and its IMEx deposition, and the experiments were performed in K562 cells, which carry BCR::ABL1 and no KMT2A rearrangement.

    PMID 39251607DOI
    What this evidence cannot show

    Full text read from PMC11385195, body element present, 141,821 characters; string search was case-sensitive on the exact symbols and would not catch an alias spelling such as AF4 in a figure label.

    Evidence that argues against this (1)
    • A paper not naming a prey is the normal state of a large-scale interactome study and is not a defect in that paper; the defect is in an auditing method that expects to find preys named in abstracts. Göös H et al. 2022
Connected to