Oncofusion

start / graph / ana-none-of-the-eight-unwritten-partners-has-an-experimental-nuclear-localisation

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q74-interactome-gap

Zero of eight unwritten AFF4/MEN1 partners has an experimentally coded nuclear localisation, and the two that are called nuclear are called it by inference

also called ECO:0000269 · ECO:0000250 · ECO:0000305 · compartment mismatch · PBX2 · AP2B1 · GOLGA2 · MTUS2 · TRAF2 · JMY · LRRK2 · RPRM

The proteins at the centre of this disease live in the nucleus, and every one of them has an experiment saying so. Of the eight database-listed partners this beat examined, not one has an experiment placing it in the nucleus - two are called nuclear, but only by guesswork from related proteins.

The same thing, for a scientist

Every subcellular-location comment on all eight partners and all three baits was pulled from UniProt with **evidence codes read**, as brief rule 91 requires. The result is uniform and was not anticipated: **KMT2A, MEN1 and AFF4 all carry `Nucleus` with ECO:0000269. Zero of the eight partners does** [E1]. Two are called nuclear and neither call is experimental - **PBX2 is ECO:0000305 (curator inference)** and **JMY is ECO:0000250 (by similarity)** - and JMY's *cytoplasmic* calls are the ECO:0000269 ones [E1]. The other six carry no nuclear annotation at all: AP2B1 at the cell membrane and coated pit, GOLGA2 at the cis-Golgi, ERGIC and spindle pole, MTUS2 in the cytoskeleton, TRAF2 in the cytoplasm, LRRK2 across **twelve** curated locations of which none is nuclear, and **RPRM with no experimentally coded location of any kind** - its cytoplasm call is ECO:0000250 and its membrane call ECO:0000305 [E1]. **This bears directly on the assays that produced the rows.** Yeast two-hybrid reconstitutes a transcription factor in the yeast nucleus, so both constructs are co-compartmentalised by the design of the assay; a protein array (LRRK2) has no compartment at all; and AP-MS lyses the cell, mixing compartments before the pulldown. **Every assay class in this beat is blind to the one property that most cleanly separates a plausible partner from an implausible one** [E2]. **What I am not claiming.** I did not measure the background rate of experimentally coded nuclear annotation in the human proteome, so this is not a test against a null. By the rule of three, 0/8 bounds the true rate below **37.5%** at 95% confidence, which is a real bound and a weak one; and if the background were 40% the probability of 0/8 would be 1.7%, at 25% it would be 10% [E3]. **The claim is descriptive and it is still useful**, because it is uniform, because it was computed the same way for the baits as a positive control, and because it gives every entry in this beat a cited localisation record it would otherwise lack.

Go deeper
The table
proteinrolenuclear annotationcodeother curated locations
KMT2AbaitNucleusECO:0000269plus two uncoded fragment comments, N320 and C180
MEN1baitNucleusECO:0000269-
AFF4baitNucleus, ChromosomeECO:0000269-
PBX2partnerNucleusECO:0000305none
JMYpartnerNucleusECO:0000250cytoplasmic vesicle, cytoskeleton, autophagosome membrane (all ECO:0000269)
RPRMpartnernone-Cytoplasm ECO:0000250, Membrane ECO:0000305 - no ECO:0000269 anywhere
AP2B1partnernone-cell membrane, coated pit (ECO:0000269)
GOLGA2partnernone-cis-Golgi network membrane, ERGIC membrane, spindle pole
MTUS2partnernone-cytoplasm, cytoskeleton (ECO:0000269)
TRAF2partnernone-cytoplasm (ECO:0000269) - one location only
LRRK2partnernone-twelve locations: vesicle, perikaryon, Golgi, axon, dendrite, ER, synaptic vesicle, endosome, lysosome, mitochondrial outer membrane, cytoskeleton, phagosome

The KMT2A row also reproduces the observation behind brief rule 91: its two taspase1 cleavage-product location comments are uncoded, so a filter on ECO:0000269 keeps the least specific of its three location comments and discards both fragment lines.

Why every assay in this beat is blind to it

This is the reason the pattern is not visible from the interaction records themselves.

  • Yeast two-hybrid (AP2B1, GOLGA2, MTUS2, TRAF2) works by reconstituting a split transcription factor in the yeast nucleus. Both fusions are delivered there by construction. A human cytoplasmic protein and a human nuclear protein are co-compartmentalised by the assay, and a compartment mismatch is structurally undetectable.
  • Protein microarray (LRRK2) has host organism taxid:-1 (in vitro). There is no cell and therefore no compartment.
  • AP-MS (PBX2, JMY, RPRM) lyses the cell before the pulldown. Nuclear and cytoplasmic proteins are mixed in the lysate, and post-lysis association is a documented artefact class of the method.

So the compartment argument is genuinely additional information, not a restatement of something the row already encodes - which is why it is worth computing even though it is one field lookup per protein.

The limit, stated once and applying to every entry in this beat: an annotation's silence is not a measured absence. Cytoplasm ECO:0000269 means a curator read an experiment showing cytoplasmic localisation. It does not mean anyone looked for a nuclear pool and failed to find one. A minor nuclear fraction of a predominantly cytoplasmic protein is exactly the thing localisation curation under-reports, and for TRAF2 and LRRK2 - both with growing immune-cell literatures annotated mostly from other tissues - that gap is real.

The statistics I did not do, stated plainly

Brief rule 84 requires the null before the proportion, and I do not have it. I did not compute what fraction of reviewed human proteins carry an ECO:0000269 nuclear location comment. Estimates for the nuclear proteome run wide and the relevant denominator here is not "nuclear proteins" but "proteins with an experimentally coded nuclear annotation", which is smaller and which I did not measure.

What can be said without it:

  • Rule of three on a zero: 0 of 8 bounds the true rate below 3/8 = 37.5% with 95% confidence. That is an honest bound and a weak one at n = 8.
  • Sensitivity rather than a test: if the background rate were p, the probability of observing 0 of 8 is (1-p)^8. At p = 0.40 that is 1.7%; at p = 0.25, 10%; at p = 0.15, 27%. The result is suggestive at a high background and unremarkable at a low one, and which regime applies was not established.
  • The internal positive control did fire: the three baits were pulled the same way, in the same call, and all three returned ECO:0000269 Nucleus. So the instrument reads the field correctly and can return the value being reported absent.

One computation would settle it and is one pass over a UniProt download: the proportion of reviewed human entries carrying SUBCELLULAR LOCATION: Nucleus with ECO:0000269. It is in the open questions.

What this does for the graph beyond this beat

check-pi-dimensions.py reports localisation coverage at 41.1% of 314 gene entries by field, and the PI has asked for localisation as a standing saturation priority. Every entry this beat writes carries a localization field with its UniProt source, its evidence code, and the explicit statement when the code is not experimental.

The evidence code is the part that is new. A localisation field that says Nucleus and a localisation field that says Nucleus (ECO:0000305, curator inference) currently count identically in that reporter, and PBX2 and RPRM are worked examples of the difference: one of the eleven proteins audited here has no experimentally coded localisation at all, and the reporter has no way to say so.

That is the same shape as rule 89 - a coverage metric that does not require the thing it is named after - and the fix is the same: record the code beside the value, then decide what counts. Proposed, not implemented, and in the memo.

Facts and figures
scriptnone - UniProt REST per-accession fetches with cc_function and cc_subcellular_location, codes read from the evidences arrays
statusrun
result summary0 of 8 partners carries an ECO:0000269 nuclear location comment; 3 of 3 baits do. Two partners are called nuclear by non-experimental codes (PBX2 ECO:0000305, JMY ECO:0000250) and RPRM has no experimentally coded location at all.
denominator8
numerator experimental nuclear0
numerator any nuclear call2
positive control3 of 3 baits returned ECO:0000269 Nucleus through the identical code path
null not computedthe proportion of reviewed human entries carrying an ECO:0000269 nuclear location comment was NOT measured
rule of three boundtrue rate below 3/8 = 37.5% at 95% confidence
selection effectthe eight were selected as partners of nuclear baits, which should enrich for nuclear proteins and makes the zero more surprising, not less; not quantified
proposal for check pi dimensionsrecord the evidence code beside every localization value and report ECO:0000269-backed coverage separately from any-annotation coverage. Two of the eleven proteins audited here would change category, and one (RPRM) has no experimental localisation at all while counting as covered under a field-presence test.
inputs
[
  "UniProt REST JSON for Q9Y5J1, P40425, Q9NS64, Q8N9B5, Q5S007, P63010, Q08379, Q5JR59, Q12933 (partners) and Q03164, O00255, Q9UHB7 (baits), retrieved 2026-09-21"
]
table
{
  "KMT2A": {
    "role": "bait",
    "nuclear": true,
    "code": "ECO:0000269",
    "note": "plus two uncoded cleavage-product location comments (N320, C180)"
  },
  "MEN1": {
    "role": "bait",
    "nuclear": true,
    "code": "ECO:0000269"
  },
  "AFF4": {
    "role": "bait",
    "nuclear": true,
    "code": "ECO:0000269",
    "note": "also Chromosome, ECO:0000269"
  },
  "PBX2": {
    "role": "partner",
    "nuclear": true,
    "code": "ECO:0000305",
    "experimental": false
  },
  "JMY": {
    "role": "partner",
    "nuclear": true,
    "code": "ECO:0000250",
    "experimental": false,
    "note": "all cytoplasmic calls are ECO:0000269"
  },
  "RPRM": {
    "role": "partner",
    "nuclear": false,
    "code": null,
    "note": "Cytoplasm ECO:0000250, Membrane ECO:0000305 - NO ECO:0000269 anywhere on the record"
  },
  "AP2B1": {
    "role": "partner",
    "nuclear": false,
    "locations": [
      "cell membrane",
      "membrane coated pit"
    ]
  },
  "GOLGA2": {
    "role": "partner",
    "nuclear": false,
    "locations": [
      "cis-Golgi network membrane",
      "ERGIC membrane",
      "spindle pole"
    ]
  },
  "MTUS2": {
    "role": "partner",
    "nuclear": false,
    "locations": [
      "cytoplasm",
      "cytoskeleton"
    ]
  },
  "TRAF2": {
    "role": "partner",
    "nuclear": false,
    "locations": [
      "cytoplasm"
    ]
  },
  "LRRK2": {
    "role": "partner",
    "nuclear": false,
    "n_locations": 12
  }
}
sensitivity
{
  "p_0.40": 0.017,
  "p_0.25": 0.1,
  "p_0.15": 0.272,
  "formula": "(1-p)^8"
}
assay blindness
{
  "yeast two-hybrid": "both fusions delivered to the yeast nucleus by construction (AP2B1, GOLGA2, MTUS2, TRAF2)",
  "protein array": "host organism taxid:-1, no cell and no compartment (LRRK2)",
  "AP-MS": "cell lysed before pulldown, compartments mixed (PBX2, JMY, RPRM)"
}
How this could bite us back
  • Used carelessly, this becomes a filter that discards every interaction between a nuclear protein and a protein annotated elsewhere - which would reject real regulated-translocation biology, a category this disease contains (the CRM1 export axis, the PICALM nuclear pool, the taspase1 cleavage fragments). The argument is a prior for downgrading unreplicated rows, not a rule for excluding partners.
  • It also cuts at this graph's own entries. If an ECO:0000269 nuclear annotation is the bar, several proteins already written here would not clear it, and applying the standard only to new partners would be exactly the asymmetry this project keeps finding in its own instruments.
Still open
  1. What fraction of reviewed human UniProt entries carries an ECO:0000269 nuclear location comment? One pass over a UniProt download turns this description into a test, and without it the 0/8 cannot be called surprising.
  2. Should check-pi-dimensions.py grade localisation by evidence code? RPRM would count as covered under a field-presence test while having no experimental localisation of any kind, which is the rule-89 shape in a new dimension.
  3. Do TRAF2 and LRRK2 have nuclear pools in haematopoietic cells that the neuronal- and immunology-weighted annotation has not captured? Two fractionation blots would convert the weakest argument in this beat into a measurement.
Every source behind this entry (3)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    Of eight IntAct-listed partners of AFF4 and MEN1 with no entry in this graph, none carries an ECO:0000269 nuclear subcellular-location comment in UniProt; PBX2's nuclear call is ECO:0000305 and JMY's is ECO:0000250, RPRM carries no experimentally coded location at all, and the three baits KMT2A, MEN1 and AFF4 all carry Nucleus with ECO:0000269.

    What this evidence cannot show

    A single snapshot of one database on one date. The cited experimental papers behind each ECO:0000269 were not opened, so per brief rule 91 the code establishes that a curator read an experiment, not that the experiment is about the compartment question asked here. Absence of a nuclear comment is an annotation's silence, not a measured absence of nuclear protein. TRAF2 and LRRK2 in particular have immune-cell literatures annotated largely from other tissues.

    Evidence that argues against this (1)
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    Every assay class producing the rows in this beat is structurally blind to subcellular compartment: yeast two-hybrid reconstitutes both fusions in the yeast nucleus, the LRRK2 row is a protein microarray with host organism taxid:-1 (in vitro), and AP-MS lyses the cell before the pulldown.

    What this evidence cannot show

    This is a statement about what the assays can detect, drawn from the recorded method and host fields, not from a study measuring false-positive rates by compartment. Post-lysis association in AP-MS is a documented artefact class but its rate in these specific datasets was not established here. Some AP-MS protocols fractionate before lysis and would not have this exposure; the BioPlex and Ewing protocols were not read in that detail.

    Evidence that argues against this (1)
    • Compartment blindness is also what makes these assays useful for discovering interactions that occur only transiently or in a small subpopulation of cells. A method that could only report interactions between proteins with matching steady-state annotations would recover nothing new. Rolland T et al. 2014
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    No background rate was computed for experimentally coded nuclear annotation in the human proteome, so 0 of 8 is bounded by the rule of three at below 37.5% with 95% confidence, and the probability of the observation is 1.7% if the background rate were 40%, 10% if 25% and 27% if 15%.

    What this evidence cannot show

    The (1-p)^8 sensitivity assumes the eight proteins are independent draws from the proteome, which they are not - they were selected by being partners of nuclear baits, which if anything should ENRICH for nuclear proteins and makes the zero more surprising rather than less. That selection effect is stated and not quantified. The rule of three assumes independence too. The honest summary is that this is a description with a bound attached, not a test.

    Evidence that argues against this (1)
    • This project has four cycles of history in which a proportion reported without its null was aggregated into a pattern across beats. Anyone quoting '0 of 8' onward should quote the 37.5% bound with it, because the point estimate will travel and the interval will not. Oncofusion knowledge graph, docs/QUEEN_BRIEF.md rule 84 2026
Connected to