Oncofusion

start / graph / ana-aff1-interactome-gap

Analysisunrated confidence 0 of 5: unrated

The curated interaction database knows less about AFF1 than this graph does

also called AFF1 interactome gap · IntAct cross-check · P51825 interactors

We checked our list of AFF1's binding partners against the big public database of protein interactions, expecting to find things we had missed. The database knows of 25; we already have 126 entries.

The same thing, for a scientist

IntAct returns **91 interaction rows and 25 distinct human partners** for AFF1 (P51825). The interactome beat has **126 entries**. Symbol-level overlap is 12 of 25 (48.0%), an upper bound. **Eleven of the thirteen absentees are single-paper, single-method anti-tag co-immunoprecipitation** with a cast (MYH7, SLC16A6, PLCD3, FGF12, CAMKV) that reads as one large-scale tagged-ORF pulldown. Only HDAC1, ZRANB2 and SBDS are worth following. This retires 'cross-check the interactome against a curated database' as a route to missing biology for this protein.

Go deeper
The thirteen absentees
genepapersrowsmethod
CAMKV22anti tag coimmunoprecipitation
PIP4K2A22anti tag coimmunoprecipitation
EPB41L111anti tag coimmunoprecipitation
EPB41L311anti tag coimmunoprecipitation
FGF1211anti tag coimmunoprecipitation
HDAC111anti tag coimmunoprecipitation
MRPL1211cross-linking study
MYH711anti tag coimmunoprecipitation
PLCD311anti tag coimmunoprecipitation
SBDS11proximity-dependent biotin identification
SLC16A611anti tag coimmunoprecipitation
ZNF52611anti tag coimmunoprecipitation
ZRANB211proximity-dependent biotin identification

The method column is the finding. Anti-tag co-immunoprecipitation, one paper each, for a set of proteins with no plausible shared compartment with a nuclear elongation factor — cardiac myosin, a monocarboxylate transporter, a phospholipase, a fibroblast growth factor. That is one screen's output, not thirteen findings.

HDAC1 (PMID 26496610), ZRANB2 and SBDS (both BioID, PMID 39251607) are the three worth an hour. ZRANB2 is an RNA-binding splicing regulator, which is the right compartment and the right process for an SEC subunit, and the interactome beat's known blind spot is precisely post-transcriptional regulators — the same class that hid IGF2BP3.

The filter that removes noise also removes HIV-1 Tat

Restricting to human-human pairs is necessary: the first row IntAct returns for AFF1 pairs it with a Francisella tularensis protein from a two-hybrid pooling approach. The same filter removes HIV-1 Tat.

That is not noise. Tat recruitment of the super elongation complex is how AFF1/AFF4 and P-TEFb were characterised in the first place, and it is the canonical case of the SEC being hijacked by something that is not KMT2A — which is, structurally, the same question this project asks about the fusion. A 'human only' filter on an interaction database silently deletes viral interactions, and for an elongation factor those are among the best-mapped in the literature.

Genuinely excluded as non-human: MAD2L1BP, RCC1, RRBP1, TAT.

A defect in this script, caught before it was written up

The first version emitted nonhuman_partners_dropped, listing every gene with at least one non-human row. MLLT1/ENL was in it. The conclusion 'the taxonomy filter drops ENL, a central AFF1 partner' was drafted as a methodological finding and was wrong: MLLT1 has 9 human rows and 2 non-human, so it was never excluded. The field conflated 'had some non-human rows' with 'was dropped'.

It is now split into partners_excluded_as_nonhuman and partners_with_some_nonhuman_rows. Same shape as everything else this cycle: a name that meant something narrower than it said, producing a clean and plausible wrong conclusion.

Still open
  1. Read PMID 39251607 for ZRANB2 and SBDS. BioID in the right compartment, and RNA-binding regulators are this beat's documented blind spot — the class that hid IGF2BP3.
  2. Read PMID 26496610 for the AFF1-HDAC1 co-IP. A histone deacetylase contacting an elongation factor is mechanistically interesting whichever way it resolves.
  3. Identify the tagged-ORF screen behind the eleven single-method hits and judge it as one source rather than eleven, which is what the load-bearing-citations audit would require.
  4. Repeat this against BioGRID and STRING before concluding the curated record is thin. One database's coverage is one database's curation priorities.
  5. Do the same diff for the KMT2A side (Q03164) and for MEN1, where the curated record may be richer because those proteins have had curation drives.
Every source behind this entry (3)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    IntAct returns 91 interaction rows for AFF1 (UniProt P51825), resolving to 25 distinct human interaction partners. This graph's interactome beat holds 126 entries. 12 of the 25 partner symbols appear somewhere in the graph's ids, titles or aliases (48.0%).

    What this evidence cannot show

    Queried 2026-09-19 via PSICQUIC, MITAB2.5, maxResults 2000; 91 rows returned, so the result is not truncated. A symbol counts as 'in graph' if it appears in any node id, title or alias, which includes passing mentions in enumeration entries — so 48% is an upper bound on real coverage and the 13-gene gap a lower bound. IntAct is one database: BioGRID, STRING and the IMEx consortium members were not queried, and a partner curated only elsewhere would not appear here. **Curation drift, noted 2026-09-20:** the same query now returns 89 interaction rows rather than the 91 recorded here. IntAct is re-curated continuously, so any row count from it is a reading on a date and not a constant. The partner counts and the argument are unaffected.

    Evidence that argues against this (3)
    • The comparison flatters the graph by counting different things. IntAct curates **direct physical interaction evidence** for AFF1 specifically; the interactome beat's 126 entries include complex subunits inferred from complex membership, proteins tied to the fusion rather than to AFF1, and regulators that never touch it. A like-for-like count of entries asserting a direct AFF1 physical interaction would be far below 126, and might not exceed 25. Oncofusion knowledge graph (internal analysis) 2026
    • A curated database being small is partly a statement about AFF1's curation priority, not about the literature. IntAct curates on request and by project, so a protein outside a curation drive is thinly covered regardless of how much has been published — which means 'the database is poorer than the graph' may describe IntAct's workload rather than this project's thoroughness. EMBL-EBI IntAct 2026
    • A second beat re-ran this query nine days later and got a different row count, which is a useful reminder about every database number in this graph: they are timestamps, not facts. The same beat also showed the query used here was the wrong instrument for the question - a symbol-driven IntAct search cannot see a protein recovered only as a PREY in a proximity deposition, which is exactly how AFF1 appears in the 2024 RBP map. EMBL-EBI IntAct 2026
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    Eleven of the thirteen absent partners rest on a single publication and a single method, predominantly anti-tag co-immunoprecipitation, and include MYH7, SLC16A6, PLCD3, FGF12 and CAMKV — proteins with no shared compartment or process with a nuclear elongation factor. HDAC1, ZRANB2 and SBDS are the exceptions worth pursuing, ZRANB2 being an RNA-binding splicing regulator.

    What this evidence cannot show

    'Implausible' is a judgement from compartment and function, not a measurement, and it is the kind of judgement that has been wrong here before — IGF2BP3 entered this graph precisely because a plausibility-driven enumeration had missed it. The source publications were not read; the method and publication counts come from the IntAct record.

    Evidence that argues against this (2)
    • Dismissing single-method single-paper hits as screen artefact is exactly the reasoning that keeps an interactome beat inside what it already believes. A large-scale tagged-ORF screen is low-precision, not zero-precision, and the three kept for follow-up were selected by the same prior that discarded the other ten. The honest position is that none of the thirteen has been evaluated. Oncofusion knowledge graph (internal analysis) 2026
    • Over-expression co-IP has a specific, well-known failure mode that cuts the other way too: tagged bait at non-physiological abundance recovers abundant structural proteins. MYH7's presence is as likely to indicate that the screen's controls were weak as that this particular row is wrong, which would put the whole screen's contribution to this list in doubt rather than just its implausible members. EMBL-EBI IntAct 2026
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    Restricting the partner list to human-human pairs removes four entities that have no human interaction row: MAD2L1BP, RCC1, RRBP1 and HIV-1 Tat. Tat's recruitment of the super elongation complex is the context in which AFF1/AFF4 and P-TEFb were characterised, so a taxonomy filter applied to an elongation factor removes one of its best-mapped interactions.

    What this evidence cannot show

    The exclusion is a property of this script's filter, not of IntAct, which curates the Tat interaction normally. MAD2L1BP, RCC1 and RRBP1 appearing only with non-human taxids suggests ortholog records rather than a judgement that the human interactions do not exist, and they were not investigated.

    Evidence that argues against this (1)
    • The filter is still right for its purpose. Without it the list is led by a *Francisella tularensis* protein from pooled two-hybrid, and pooled Y2H cross-species hits are the highest-false-positive category in the whole database. Losing one well-known viral interaction that this graph already covers elsewhere is a cheap price. Oncofusion knowledge graph (internal analysis) 2026
Connected to