Oncofusion

start / graph / interface-aff4-fmr1-declined-an-alzheimers-data-integration-paper-curated-as-an-experiment

Interfaceweak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.q160-the-prey-only-interactome

AFF4-FMR1 - DECLINED, and the provenance is the finding: the 'publication' behind this IntAct row is HENA, an Alzheimer's data-integration dataset, not an experiment on AFF4

also called AFF4-FMR1 · AFF4-FMRP · HENA

A database lists FMRP, the fragile-X protein, as an AFF4 partner. The paper it points to is not an experiment at all - it is a project that merged other people's databases into one big network for Alzheimer's research.

The same thing, for a scientist

**Verdict: DECLINED, and the reason generalises beyond this pair.** IntAct holds one AFF4-FMR1 record: **2h fragment pooling**, *physical association*, host ***Saccharomyces cerevisiae***, AFF4 as prey, intact-miscore 0.37, from **PMID 31413325** [E1]. That PMID is **HENA** (Sugis et al., *Scientific Data* 2019), described by its own authors as *a heterogeneous network-based data set for Alzheimer's disease* assembled by combining results stored in databases designed for different data types, and demonstrated with graph convolutional networks [E2]. PubMed types it **Dataset**. Its author list includes Hybrigenics staff, so the underlying two-hybrid screens are real - **but the cited object is a data-integration resource, and a record whose provenance is 'a dataset that aggregated other datasets' is one curation layer further from an experiment than any other row in this census.** The AFF-family angle that would otherwise make this interesting - FMR1's paralogue **AFR2/AFF2 (FMR2)** is a member of the same AFF family as AFF1 and AFF4 - is a reason to look harder, not a reason to believe this row.

Go deeper
Three separate reasons, any one of which is sufficient
  1. The host organism is yeast. 2h fragment pooling in S. cerevisiae tests whether two fragments can reconstitute a transcription factor in a yeast nucleus. It says nothing about a human nucleus, and this corpus already records that the four AFF4 two-hybrid partners are one S. cerevisiae screen split across three PSI-MI labels.
  2. The bait was an Alzheimer's-relevant protein, not AFF4. AFF4 is the prey.
  3. The publication is a dataset, not a study. A reader following the citation reaches a resource description, not an experiment, and cannot find the original screen conditions from it.

The general point worth recording: interaction databases curate publications, and a publication can be a database. When a row's PMID resolves to a Dataset article type, the chain of custody from observation to record has an extra, undocumented hop in it. Nothing in a MITAB row or an IntAct REST record exposes this. The only way to see it is to fetch the PMID and read what the article actually is - which this beat did, and which is the reason two of its four strongest-looking leads collapsed.

The AFF-family connection, which is real and is not this row

FMR1 encodes FMRP and is silenced in fragile-X syndrome. AFF2, also called FMR2, sits in the same AFF family as AFF1 and AFF4 and is silenced in FRAXE intellectual disability. The two genes are neighbours in the same region of Xq27-28 and were cloned within a few years of each other for the same reason.

So AFF-family protein interacts with FMRP is a hypothesis with an independent motivation: the AFF family and the fragile-X locus have a shared history, and AFF4's scaffold function in the super elongation complex is exactly the kind of role FMRP-associated translational control intersects with.

This entry does not make that argument on the basis of this row. A yeast two-hybrid fragment pair curated out of an Alzheimer's data-integration resource is not evidence for it. The argument is recorded here so that the next beat that encounters AFF4-FMR1 knows both that the row is bad and that the underlying question has a rationale.

Facts and figures
partner agene-aff4
partner bFMR1
partner b uniprotQ06787
partner b length AA632
partner a uniprotQ9UHB7
partner a length AA1163
verdictDECLINED
adjudicated byq160, cycle 146, 2026-09-24
druggabilityNot assessable.
n methods1
n platforms1
n publications1
endogenous sideno - heterologous two-hybrid in yeast
intact miscore0.37
cell statenot applicable - heterologous reconstitution
epigeneticsNone.
decline reason classthe cited publication is a data-integration resource (PubMed article type: Dataset), not an experiment - one curation layer further from observation than any other row in this census
absence classnever-measured
absence class noteNo human-cell experiment has tested AFF4 against FMRP. Given the AFF2/FMR2 family history the question is not unreasonable, but no reagent-level work exists.
interface residues a
[]
structures
[]
ligands
[]
records
[
  "AFF4-FMR1, 2h fragment pooling, physical association, S. cerevisiae, AFF4 prey, miscore 0.37, PMID 31413325"
]
cell systems
[
  "Saccharomyces cerevisiae"
]
localization
{
  "compartment": "not measured in this record",
  "subnuclear": [],
  "shown_for_fusion": "NO - nobody has looked in any fusion."
}
measurements
[
  {
    "quantity": "count",
    "subject": "rows in this census whose PMID resolves to a PubMed article typed Dataset",
    "value": 1,
    "units": "rows",
    "n": 212,
    "universe": "the 2026-09-24 prey-only census; article types were checked for the 20 shortlisted PMIDs only, not all of them",
    "null": null,
    "note": "1 of 20 PMIDs actually checked. The denominator is the shortlist, NOT the full census, and the true rate across all 212 is unmeasured."
  }
]
How this could bite us back
  • Recording 'the citation is a dataset' as a decline reason could be abused to dismiss legitimate primary depositions that happen to be published as data descriptors. The criterion should be whether the original experimental conditions are recoverable from the cited object, not what PubMed calls it.
Still open
  1. How many IntAct rows across this whole graph cite a PubMed record typed Dataset rather than Journal Article? Only 20 PMIDs were checked here and one was a dataset; the corpus-wide rate is unknown and is cheap to measure.
  2. Does AFF4 or AFF1 associate with FMRP in a human cell? The AFF2/FMR2 family history gives the question a rationale that this row does not supply.
Every source behind this entry (2)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    IntAct REST returns one AFF4-FMR1 (Q9UHB7-Q06787) record: detection method '2h fragment pooling', interaction type 'physical association', host organism Saccharomyces cerevisiae (Baker's yeast), AFF4 as PREY, intact-miscore 0.37, from PMID 31413325.

    What this evidence cannot show

    One record. The host-organism field records where the assay ran (yeast), not the species of the proteins, both of which are human here. This corpus has previously been caught treating three PSI-MI labels from one yeast screen as three experiments.

    Evidence that argues against this (1)
    • Hybrigenics two-hybrid screens are systematic and well-controlled, and their fragment-level resolution is a genuine strength - a fragment-pooling hit identifies the interacting region. The objection is to the citation's provenance and the heterologous system, not to two-hybrid as a method. Sugis E et al. 2019
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    PMID 31413325 (Sugis et al., Scientific Data 2019) describes HENA, a heterogeneous network-based data set for Alzheimer's disease, assembled by combining the results of experiments stored in separate databases in order to give a systematic view, and demonstrates graph convolutional network analysis on it. PubMed types the article as Dataset. Neither AFF4 nor the super elongation complex appears in the title or abstract.

    HENA, heterogeneous network-based data set for Alzheimer's disease. · Sugis E, Dauvillier J, Leontjeva A, Adler P, Hindie V, Moncion T, Collura V, Daudin R, Loe-Mie Y, Herault Y, Lambert JC, Hermjakob H, Pupko T, Rain JC, Xenarios I, Vilo J, Simonneau M, Peterson H · Sci Data 2019
    PMID 31413325DOI
    What this evidence cannot show

    Abstract and article-type read; the HENA resource itself was not downloaded, so the original provenance of the AFF4-FMR1 pair inside it - which screen, which fragment, which laboratory - was not traced. The author list includes four Hybrigenics staff and Henning Hermjakob of EMBL-EBI, so the two-hybrid data are real primary data that HENA aggregated; the objection is that the cited record points at the aggregate.

    Evidence that argues against this (1)
    • A data-integration paper can be the correct citation for data it publishes for the first time, and the two-hybrid screens contributed by Hybrigenics may appear nowhere else. In that case IntAct's choice of PMID is the only one available and the row's provenance is no worse than any other unpublished-screen deposition. Sugis E et al. 2019
Connected to