Oncofusion

start / graph / concept-what-a-curated-interaction-row-actually-claims

Explainerestablished confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.Q76-spoke-provenance

What a curated interaction row actually claims: spoke expansion, and the ladder from 'was in the same tube' to 'touches'

also called spoke expansion explained · MI:1060 · association versus physical association versus direct interaction · why 'X binds Y' in a database may mean neither

Databases store protein partnerships as simple pairs - A binds B - but most of those pairs come from one experiment that fished out a whole crowd at once, and the pairing was written down afterwards by a curator rather than measured. The database says so, in a field almost nobody reads.

The same thing, for a scientist

A pulldown is an n-ary experiment: one tagged bait, many co-purifying preys. To store it as binary pairs a curator applies the **spoke model** - each prey is paired with the bait and nothing is asserted between preys - and records the transformation as `psi-mi:MI:1060 (spoke expansion)` in the expansion field [E1]. The alternative, `MI:1061 matrix expansion`, pairs every molecule with every other and is weaker still. Orthogonally, the interaction-type field carries a three-rung ladder: **MI:0914 association** (may be in one *or more* complexes - typically what a spoke row gets), **MI:0915 physical association** (in the same physical complex), **MI:0407 direct interaction** (in direct contact) [E1]. `MI:2364 proximity`, the BioID/TurboID class, is a fourth thing on neither ladder: nanometre co-occurrence, which is not a complex and not a contact [E1]. **Reading MI:1060 as 'nobody observed this' is wrong**, and this project got it wrong once before catching it: IntAct spoke-expands an n-ary deposition whatever produced it, so a ternary crystal structure comes out flagged too - PSIP1-KMT2A from PMID 22327296 is flagged MI:1060 *and* typed MI:0407 by x-ray diffraction [E2][E3]. The correct test is the conjunction of the expansion flag with the type and the detection method. Measured across this project's six core proteins, 63.9% of records and 70.6% of pairs are spoke-expanded [E4].

Go deeper
The picture, in one paragraph

Imagine fishing a labelled protein out of a cell with a magnet. Everything stuck to it comes up too - directly attached, attached via something else, or just along for the ride. You now know the crowd, not the handshakes.

To put that in a database of pairs, a curator writes down bait-with-prey1, bait-with-prey2, and so on: spokes radiating from a hub. That is a faithful, conservative summary of the experiment - it does not invent prey-to-prey links - but every one of those spokes is an inference. The prey might touch the bait; it might be three proteins away.

The honest thing about IntAct is that it says so on every row. The dishonest thing is what happens next: the row reaches a downstream user as A interacts with B, and the flag does not travel with it.

The two independent fields, which get confused

Expansion answers how was this binary row produced?

termmeaning (read back from the ontology)
(empty)not expanded - the record as deposited was already binary
MI:1060 spoke expansionn-ary data expanded assuming all molecules interact with a single designated molecule, usually the bait
MI:1061 matrix expansionn-ary data expanded assuming all molecules interact with each other

An empty expansion field is the NEGATIVE, not a missing value. In MITAB 2.7 it is -; in the IntAct REST representation the expansionMethod key is null. Either way it means this was not expanded, and treating it as unknown inflates uncertainty in the wrong direction.

Interaction type answers what is being claimed about the two molecules?

termmeaning
MI:0914 associationmay participate in formation of one, but possibly more, physical complexes
MI:0915 physical associationwithin the same physical complex
MI:0407 direct interactionin direct contact with each other
MI:2364 proximitycoincident occurrence within nanometres, by proximity labelling
MI:0403 colocalizationcoincident in a subcellular fraction; explicitly says a physical interaction cannot be inferred

The two fields are independent, and the combination is the information. Spoke-expanded + MI:0914 + anti-tag co-IP is the weakest common cell and is what most atlas rows are. Not expanded + MI:0407 + ITC is the strongest.

The trap: a crystal structure flagged as an inference

A ternary structure is n-ary data. A curator storing menin/MLL/LEDGF as binary rows applies the spoke model, correctly, and the resulting rows carry MI:1060 - including the row typed MI:0407 direct interaction with detection method x-ray diffraction 2.

So MI:1060 alone is not a grade of evidence quality. It is a statement about a representation. The test that means nobody observed this binary pair is a conjunction:

every record is spoke-expanded, AND no record is typed MI:0407, AND no record's detection method is structural (x-ray, NMR, EM) or a binary biophysical method (ITC, SPR, fluorescence polarisation, CD, sedimentation).

Applied to this project's own written interactions, the naive flag-only test misclassified 1 of 27 4. That is a low rate, and the reason it is low is instructive: this classifier reads a stored field rather than inferring a latent property, which is the kind of classifier worth building.

Why this matters more here than in most graphs

This project's central therapeutic argument is about a pocket. A small molecule occupies the site on menin that the MLL MBM motifs insert into; the whole programme depends on that being a direct contact with a defined interface, not a co-membership.

A spoke-expanded row cannot support an interface claim. It can support a complex-membership claim, and complex membership is real and often what an entry actually needs - the WRAD subunits are in the MLL1 complex and the graph is right to say so. The error to avoid is silently upgrading one to the other, which is what happens when MI:0914 becomes the word "binds" in a sentence.

The practical rule for this graph:

  • Use binds only when a binary observation exists, or say in the label that it is a complex-membership claim.
  • Record the provenance grade in the evidence caveat, naming the expansion flag, the interaction type and the detection method - so a later reader can see the ladder without re-querying.
  • A spoke row is not worthless. The prey was in the pulldown. That is a fact about the cell and it belongs in the graph.
Facts and figures
correct test for unobservedevery record spoke-expanded AND no MI:0407 record AND no structural or binary-biophysical detection method
expansion terms
{
  "": "not expanded - deposited binary. An EMPTY field is the negative, never 'unknown'.",
  "MI:1060": "spoke expansion - every prey paired with the bait, nothing asserted prey-to-prey",
  "MI:1061": "matrix expansion - every molecule paired with every other; weaker still"
}
interaction type ladder
[
  {
    "term": "MI:0407",
    "label": "direct interaction",
    "claims": "the two molecules are in direct contact"
  },
  {
    "term": "MI:0915",
    "label": "physical association",
    "claims": "the two molecules are in the same physical complex"
  },
  {
    "term": "MI:0914",
    "label": "association",
    "claims": "the two may be in one, but possibly more, physical complexes"
  },
  {
    "term": "MI:2364",
    "label": "proximity",
    "claims": "nanometre co-occurrence by proximity labelling; neither complex nor contact"
  },
  {
    "term": "MI:0403",
    "label": "colocalization",
    "claims": "same subcellular fraction; the definition states a physical interaction cannot be inferred"
  }
]
mitab 2 7 field positions
{
  "note": "MITAB column numbers in the literature are 1-based; a tab-split array is 0-based, and this project has been confused by the off-by-one once.",
  "publication": "0-based col8 / 1-based 9",
  "interaction_type": "0-based col11 / 1-based 12",
  "interaction_ac": "0-based col13 / 1-based 14",
  "confidence_intact_miscore": "0-based col14 / 1-based 15",
  "expansion_method": "0-based col15 / 1-based 16",
  "experimental_role_A_B": "0-based col18 and col19",
  "author_confidence": "embedded in the pipe-delimited annotation string at 0-based col26; split on | and match the author-confidence: prefix rather than regexing the whole row, because isoform-comment text contains colons",
  "host_organism": "0-based col28"
}
rest equivalents
{
  "expansion": "expansionMethod (null when not expanded)",
  "type": "type and typeMIIdentifier",
  "detection": "detectionMethod and detectionMethodMIIdentifier",
  "score": "intactMiscore",
  "note": "the REST route also reaches processed-chain and isoform accessions that PSICQUIC id:<acc> does not"
}
measured here
{
  "records_spoke_pct": 63.9,
  "pairs_spoke_only_pct": 70.6,
  "naive_test_misclassification": "1 of 27"
}
How this could bite us back
  • A grading vocabulary makes deletion feel principled. The single most likely misuse of this concept is an agent removing a spoke-supported partner entry on the grounds that it is 'only' an association - when the correct action is always a caveat and a queue, because the prey really was in the pulldown and the interaction may well be real.
  • The inverse misuse is as bad: an entry that scores MI:0407 direct interaction can still be a two-hybrid artefact from a screen, and the type field will not say so. Reading the ladder without reading the detection method will confidently promote the wrong rows.
Still open
  1. Should this graph carry the expansion flag, interaction type and detection method as structured fields on every interaction edge, rather than as prose in an evidence caveat? The data exist for every pair in the census and the edge schema has room.
  2. Proximity labelling (MI:2364) is 111 of the 908 records here and fits neither ladder. Does it deserve a provenance grade of its own in this graph, given that the TurboID of KMT2A::AFF1 this project keeps proposing would generate exactly that class of evidence?
  3. How often does MI:1061 matrix expansion appear in a leukaemia-relevant interactome at all? None of the 908 records here carries it, but the query was not designed to find it and a matrix-expanded row would be the weakest thing in the vocabulary.
Every source behind this entry (4)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    The PSI-MI controlled vocabulary defines MI:1060 spoke expansion as n-ary data expanded to binary assuming all molecules interact with a single designated molecule usually the bait, MI:1061 matrix expansion as the all-against-all alternative, MI:0914 association as molecules that may participate in one but possibly more physical complexes, MI:0915 physical association as within the same physical complex, MI:0407 direct interaction as in direct contact, MI:2364 proximity as coincident occurrence within nanometres detected by proximity labelling, and MI:0403 colocalization as coincidence in a subcellular fraction from which a physical interaction cannot be inferred.

    What this evidence cannot show

    Fetched live rather than quoted from a dispatch brief, because the brief that supplied these codes has been wrong about a premise in three consecutive cycles; the premise checked out. One defect noticed and not acted on: OLS4's definition of MI:1061 matrix expansion opens 'expanded to binary using the spoke model', which is a copy-paste of the MI:1060 text, though its second sentence correctly says all molecules interact with each other. A vocabulary definition constrains intent, not twenty-one years of individual curation decisions.

    Evidence that argues against this (1)
    • The ladder implies an ordering that the terms do not strictly have. MI:0915 physical association from a stringent reconstitution of purified components can be far better evidence than MI:0407 direct interaction inferred from a single yeast two-hybrid, and the type field records what the curator concluded rather than how well it was shown. The detection method is doing more work in this concept than the type is. EMBL-EBI Ontology Lookup Service 4 2026
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    IntAct's two records for the PSIP1-KMT2A pair, both from PMID 22327296, are both flagged MI:1060 spoke expansion, and one of them (EBI-15969639) is typed MI:0407 direct interaction with detection method x-ray diffraction.

    What this evidence cannot show

    Two records on one day. Whether flagging a ternary-structure record as spoke-expanded is a curation choice or a pipeline artefact was not established here - both readings are consistent with what the records show, and under the first reading the curation is correct and the error is entirely in how the flag is read downstream, which is the position this concept takes.

    Evidence that argues against this (1)
    • If a curator deliberately spoke-expanded a three-body structure, the flag is doing its job and the entire framing of MI:1060 as a weakness marker is a downstream misuse rather than a database problem. This concept would then be about reader discipline, not about data quality - which is the more likely of the two readings and is why the corrected test is a conjunction rather than a repudiation of the flag. EMBL-EBI Ontology Lookup Service 4 2026
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    PMID 22327296 (Huang J, Gurung B, Wan B, et al., Lei M, Nature 2012) is the source of the menin/MLL/LEDGF ternary crystal structure PDB 3U88, whose IntAct curation produced the spoke-flagged x-ray-diffraction record.

    PMID 22327296DOI
    What this evidence cannot show

    Title, nine-author list, journal, year and DOI copied out of a live NCBI efetch XML record fetched 2026-09-21 using the anchored PubmedData/ArticleIdList/ArticleId path (brief rule 81). Already cited in 52 entries of this graph (brief rule 82); used here only as the identified source of a database record, with no new claim about its findings. The paper's abstract does not mention LEDGF; the ternary content was taken from the PDB 3U88 entity list.

    Evidence that argues against this (1)
    • Choosing this paper as the worked example makes the trap look rarer than it is, because it is a famous structure that any reader of this graph would recognise as observed. The dangerous cases are spoke-flagged structural records for pairs nobody in the project knows by name, where no reader's memory will catch the misclassification. RCSB Protein Data Bank 2026
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    Across the IntAct records of KMT2A, MEN1, PSIP1, AFF1, AFF4 and DOT1L fetched 2026-09-21, 580 of 908 records (63.9%) and 317 of 449 partner pairs (70.6%) are spoke-expanded; applying the flag alone as a test of 'nobody observed this' to the 27 spoke-only pairs this graph has written misclassified one.

    What this evidence cannot show

    One database, one day, six proteins - these proportions describe this domain and are not a statement about IntAct as a whole. The full census, its null and its intervals are in ana-two-thirds-of-the-curated-interactome-is-a-row-nobody-observed; this concept entry cites the headline only so a reader meeting spoke expansion for the first time has a sense of scale.

    Evidence that argues against this (1)
    • A high spoke fraction is partly a property of the era and the method rather than of the proteins: affinity purification with mass spectrometry became the dominant interaction assay after about 2005, and any protein whose interactome was mapped in that period will look spoke-heavy. The figure would have been much lower in 1998 and says as much about how the field works as about what is known. EMBL-EBI IntAct 2026
Connected to