Two thirds of the curated interactome of this project's six core proteins is spoke-expanded - and the pairs this graph chose to write are significantly better than the ones it did not

An affinity purification is one experiment with one bait and many preys. To store it as binary pairs, a curator applies the spoke model: every prey is paired with the bait, and nothing is asserted between preys. IntAct records the transformation in the expansion field as psi-mi:MI:1060 (spoke expansion), whose ontology definition read back live is "Complex n-ary data has been expanded to binary using the spoke model. This assumes that all molecules in the complex interact with a single designated molecule, usually the bait" 1.

A spoke row is a real observation. The prey was in the pulldown. What it is not is evidence that the two proteins touch, because a prey can reach the bait through any number of bridging subunits. That distinction is the whole of this project's menin-interface argument, which is about direct contacts and about a pocket a drug occupies, so the census is worth having.

The interaction-type vocabulary carries the same gradient and was also read back live 1: MI:0914 association - "may participate in formation of one, but possibly more, physical complexes"; MI:0915 physical association - "within the same physical complex"; MI:0407 direct interaction - "in direct contact with each other". A spoke-expanded row is usually typed MI:0914, which is the weakest of the three.

A graph interaction claim is a binds or part_of edge in data/graph/edges/*.json with one endpoint in the seven core ids - gene-kmt2a, gene-men1, gene-psip1, gene-aff1, gene-aff4, gene-dot1l, fusion-kmt2a-aff1 - and the other endpoint a graph entry carrying a UniProt accession. binds is this graph's physical-contact edge; part_of between two protein entries is a complex-membership claim, which is a physical-association claim, and the WRAD subunits are encoded that way. Self-pairs are excluded. Each pair is looked up in IntAct by base accession with a gene-symbol fallback, because IntAct stores IBTK under the unreviewed protein-group leader E7EPI0 while the graph stores Q9P2D0 - a trap already carrying a row in data/symbol-collisions.json.

That yields 72 pairs, of which 51 are adjudicable against IntAct and 21 have no curated IntAct row at all. The 21 are reported as their own arm and are NOT folded into a denominator, because "no IntAct row" means "not curated here", never "not observed" 2.

The rule has a visible boundary and the boundary matters. Several interactions the project reasons from are encoded through a complex node or through related_to and are therefore invisible to it - gene-ash2l reaches gene-kmt2a only by related_to, and gene-cdk9 and gene-ell have no binds or part_of edge to any core entry at all. Those pairs are classified separately in results.json under load_bearing_core, directly from IntAct, so the load-bearing question is answered on its own terms rather than by an artefact of edge vocabulary.

Fisher exact two-sided on written versus not-written: p = 0.0051. The written pairs are a subset of the 449, so the comparison is drawn against the complement rather than against the whole, which is the honest form.

Per core protein, where the interesting variation is:

MEN1's interval does not overlap AFF1's, AFF4's or DOT1L's. The protein this project proposes to drug is the best-observed of the six, and the index fusion's own 3' partner is close to the worst.

The heuristic was: every IntAct record for this pair carries MI:1060, therefore nobody observed the binary pair. Reading all 27 rows it returned found it wrong once, and wrong for a structural reason rather than a clerical one.

PSIP1-KMT2A is two records, both flagged MI:1060, and one of them is MI:0407 direct interaction determined by X-RAY DIFFRACTION, from PMID 22327296 3. IntAct spoke-expands an n-ary deposition whatever produced it, so a ternary crystal structure - here menin/MLL/LEDGF, PDB 3U88 4 - comes out of the pipeline flagged as an inference. MI:1060 is a statement about the expansion, not about whether anybody looked.

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

The heuristic's answer is kept in results.json beside the corrected one, because the size of its error is itself a result. 3.7% is a good error rate and the cycle-42 precedent was 5 of 9; the difference is that this classifier tests a stored database field rather than inferring a latent property, and that is the kind of classifier worth building.

And the PDB arm needs its own caveat. A PDB entry containing both chains is not a direct contact. 6PWV is the MLL1 core complex on a nucleosome and contains KMT2A and DPY30, but DPY30 binds ASH2L, not KMT2A 5. The shared-entry test corroborates co-membership of one observed assembly - which is a weaker claim than a contact, and is exactly what a spoke row asserts. So for KMT2A-DPY30 the spoke flag and the structure agree, and the honest reading is that both are right.

Persisted as identities and not only as a count (brief rule 90); the full per-pair record with interaction ACs, PMIDs, detection methods and miscores is in analyses/spoke-expansion-provenance-census/results.json.

* has a PDB entry containing both chains, so co-membership of an observed assembly is corroborated even though the binary contact is not.

Six of the eight DOT1L rows are the Hein 2015 spoke set that cycle 42 already worked out in full 6, and one - MLLT6/AF17 - is a genuine DotCom subunit whose contact has still never been observed binary.

1. It does not say a spoke row is worthless. Every entry touched in this beat says the opposite in its caveat, in those words. The prey was in the pulldown; that is a fact about the cell.
2. It does not bound the literature. IntAct is a curated subset, and a co-IP shown in a supporting figure of a paper about something else never reaches it. Worked example from this run: reading MEN1 through PSICQUIC, JUND is absent from the record entirely and I wrote that absence down as a finding before catching it; through the REST endpoint it is there with six records, MI:0407, ITC and x-ray diffraction 2. The absence was my instrument's, not IntAct's.
3. It does not deduplicate to experiments. A pair with one spoke row and a pair with six spoke rows from six papers are both spoke_only here. Record counts and PMIDs are stored per pair so a stricter reading can be recomputed without re-fetching.
4. It says nothing about the fusion. fusion-kmt2a-aff1 has no UniProt accession, so no pair keyed on it joins to IntAct, and separately IntAct holds fusion-protein interactors for MLL-ENL, MLL-AFF4 and MLL-AF9 but none for KMT2A::AFF1.