Oncofusion

start / graph / ana-psicquic-id-queries-miss-half-of-kmt2as-intact-record

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q76-spoke-provenance

Every IntAct number in this project is drawn through a query that misses half the record - and what it misses is the taspase1 fragments

also called PSICQUIC id: undercount · processed-chain identifiers are invisible to PSICQUIC · Q03164-PRO_0000390949

Every count this project has taken from the IntAct interaction database was fetched with a query that silently skips any record filed under a protein fragment rather than the whole protein. For MLL, whose database records are mostly filed under its two cut pieces, that means half the record has never been seen here.

The same thing, for a scientist

Both interactome analyses in this repository - `interactome-bait-versus-prey` and `directed-interactome-coverage-recheck` - read IntAct through PSICQUIC with `id:<accession>`. Measured against the IntAct REST endpoint on the same day [E1]: for KMT2A, `id:Q03164` returns 142 distinct interaction records where `findInteractions/Q03164` returns 276, and nothing is in PSICQUIC that is not in REST, so REST is a strict superset. The reach across the six core proteins is KMT2A 51.4%, MEN1 43.6%, PSIP1 43.4%, DOT1L 60.0%, AFF4 82.6%, AFF1 83.5%. The cause is that PSICQUIC's `id:` field does not match **processed-chain identifiers** - `Q03164-PRO_0000390949` (MLL-N) and `Q03164-PRO_0000390950` (MLL-C), the taspase1 cleavage fragments this graph's architecture argument is built on - nor isoform accessions such as `O00255-2` for menin. Widening to `id:Q03164*` recovers only part of it (192 MITAB rows). Two concrete consequences: the menin-KMT2A **direct interaction** record EBI-15969392 is invisible to the query because menin is stored there as `O00255-2` [E1], and reading MEN1 through PSICQUIC makes **JUND absent from menin's interactome entirely** when it is in fact present with six records including ITC and x-ray diffraction [E1]. The pair denominators differ accordingly: cycle 34's grading counted 290 partner pairs across the six baits [E2]; the REST record holds 449.

Go deeper
The measurement
baitREST interaction recordsPSICQUIC id: recordsreach
KMT2A27614251.4%
MEN11335843.6%
PSIP11134943.4%
DOT1L603660.0%
AFF416113382.6%
AFF1796683.5%

Nothing was in PSICQUIC and not in REST, for any of the six. The relation is containment, not disagreement, which is why the failure has no signature: every number PSICQUIC returns is correct, there are simply fewer of them.

Query-form controls, run on KMT2A:

id:Q03164            161 MITAB rows / 142 interaction ACs
identifier:Q03164    161
id:Q03164*           192
Q03164 (free text)   218
REST findInteractions 319 rows / 276 interaction ACs

So this is not a matter of choosing the wrong field name. Even the wildcard does not reach it, and the free-text form is still 58 records short.

Why it is worse here than it would be anywhere else

A processed chain is UniProt's identifier for a proteolytic product. Q03164-PRO_0000390949 and Q03164-PRO_0000390950 are MLL-N (MLL^N320) and MLL-C (MLL^C180), the two fragments taspase1 cuts full-length MLL into and which re-associate through the FYRN/FYRC interface. Depositors curating an MLL complex frequently - and correctly - file against the fragment that was actually in the experiment.

So the identifier the query cannot see is the one that carries this graph's central biology. The records reached only through REST include the MLL-N / MLL-C association itself (EBI-2614630, physical association, PMID 20153263), the MLL-N interactions with HOXA9, MSX2 and CDKN1B gene loci from the PAF-complex paper, and the AFF1-side spokes filed against Q2Q440.

The same shape appears on the partner side. Menin is stored in some records as isoform O00255-2, and one of those records is EBI-15969392, menin-KMT2A, MI:0407 direct interaction, PMID 22327296 - the single most load-bearing interaction claim in this entire project, invisible to the query the project has been using.

This is brief rule 91's problem one level down: not "the evidence code points at the wrong paper" but "the accession you searched is not the accession the experiment was filed under", and it is the same family as the E7EPI0 protein-group-leader row already in data/symbol-collisions.json.

The absence I nearly wrote down, which is the honest part of this entry

Working from the PSICQUIC export, I listed menin's IntAct partners and JUND was not among them. Menin-JunD is one of the best-characterised interactions in this whole domain, with a crystal structure and an ITC-measured affinity, so I wrote it down as a finding: IntAct's coverage is itself incomplete; 'no IntAct row' therefore means 'not curated', never 'not observed'.

The conclusion is correct and worth keeping. The example was not. On the REST route JUND is there, six records, MI:0407 direct interaction, detected by anti-tag co-IP, ITC, pull down and x-ray diffraction, zero of them spoke-expanded 1. The absence was my instrument's.

The general form: when an absence in a database is about to become a claim, re-pull it through a second route before writing it. An absence produced by a narrow query reads exactly like an absence in the data, and the more famous the missing item the more likely it is that the instrument, not the world, is at fault.

What this changes, and what it does not

It does not invalidate cycle 34 or cycle 41-42. Every partner those beats found is a real partner, every gap they closed was a real gap, and the UTP18/WDR5 adjudication 3 is unaffected because it rests on the full text of PMID 15199122 rather than on a row count.

It does change every denominator. Cycle 34's grading counted 290 partner pairs across the six baits 2; the REST record holds 449. Any figure of the form "k of n partners are X" drawn through PSICQUIC has an n that is too small by a factor that varies from 1.2 to 2.3 between baits, so the per-bait comparisons those analyses make are not on a common footing. The bait most affected is PSIP1, whose famous one-bait-supported-partner-out-of-thirty-six figure 2 was drawn from 43.4% of its record.

The fix is one line - swap the endpoint - and it is applied in analyses/spoke-expansion-provenance-census/run.py, which fetches both routes and prints the comparison every run so the discrepancy cannot go stale. The two older analyses are not mine to edit and the correction is filed in this cycle's memo.

Facts and figures
scriptanalyses/spoke-expansion-provenance-census/run.py
statusrun
result summaryPSICQUIC id:<acc> reaches 43-84% of IntAct's interaction records per bait and 51.4% for KMT2A; REST is a strict superset; what is missed is concentrated in processed-chain (taspase1 fragment) and isoform accessions, including the menin-KMT2A direct-interaction record and all six MEN1-JUND records.
reach by bait PCT
PSIP143.4PSIP1: 43.4MEN143.6MEN1: 43.6KMT2A51.4KMT2A: 51.4DOT1L60.0DOT1L: 60AFF482.6AFF4: 82.6AFF183.5AFF1: 83.5
records rest
DOT1L60.0DOT1L: 60AFF179.0AFF1: 79PSIP1113.0PSIP1: 113MEN1133.0MEN1: 133AFF4161.0AFF4: 161KMT2A276.0KMT2A: 276
records psicquic
DOT1L36.0DOT1L: 36PSIP149.0PSIP1: 49MEN158.0MEN1: 58AFF166.0AFF1: 66AFF4133.0AFF4: 133KMT2A142.0KMT2A: 142
inputs
[
  "IntAct REST ws/interaction/findInteractions for the six core accessions, 2026-09-21",
  "IntAct PSICQUIC search/query/id:<accession>?format=tab27, same six, same day",
  "PSICQUIC query-form controls: id:, identifier:, id:*, free text"
]
reach by bait PCT — the numbers behind the chart
{
  "KMT2A": 51.4,
  "MEN1": 43.6,
  "PSIP1": 43.4,
  "DOT1L": 60,
  "AFF4": 82.6,
  "AFF1": 83.5
}
records rest — the numbers behind the chart
{
  "KMT2A": 276,
  "MEN1": 133,
  "PSIP1": 113,
  "DOT1L": 60,
  "AFF4": 161,
  "AFF1": 79
}
records psicquic — the numbers behind the chart
{
  "KMT2A": 142,
  "MEN1": 58,
  "PSIP1": 49,
  "DOT1L": 36,
  "AFF4": 133,
  "AFF1": 66
}
invisible identifier classes
[
  "processed chains: Q03164-PRO_0000390949 (MLL-N), Q03164-PRO_0000390950 (MLL-C)",
  "isoform accessions: O00255-2 (menin isoform 2), Q8TEK3-2 (DOT1L isoform 2)",
  "fusion-protein interactors: EBI-2890930 (MLL-AF9), EBI-2620048 / EBI-2620075 (MLL-AFF4), EBI-2620068 (MLL-ENL)"
]
named records recovered only by rest
[
  "EBI-15969392 KMT2A-MEN1 MI:0407 direct interaction, PMID 22327296, menin as O00255-2",
  "EBI-2614630 MLL-N with MLL-C, physical association, PMID 20153263",
  "the six MEN1-JUND records, MI:0407, ITC and x-ray diffraction"
]
pair denominator then and now
{
  "cycle_34_psicquic": 290,
  "this_run_rest": 449
}
How this could bite us back
  • Switching to the REST endpoint will make every interactome number in this project move at once, in the direction of more partners and lower coverage percentages, and nothing will have changed in the biology. A reader seeing coverage fall from 94.3% will read it as regression. The move needs to be announced as an instrument change, with both numbers shown, or it will be misread as a loss.
  • The larger record is also a larger surface for low-quality rows. Half of what PSICQUIC was hiding is spoke-expanded atlas output, and a coverage metric that now demands entries for all of it would push the graph toward writing up exactly the partners this cycle's census grades weakest.
Still open
  1. Does the undercount change the DIRECTION of any conclusion in interactome-bait-versus-prey or directed-interactome-coverage-recheck, or only the denominators? Re-running both against REST is a day's work and the answer is not guessable, because the missing records are not a random subset.
  2. How many other databases this project queries by accession are filed against processed chains? PhosphoSitePlus, Reactome and ComplexPortal all model MLL-N and MLL-C, and no analysis here has ever asked whether its query reaches them.
  3. Should this graph carry Q03164-PRO_0000390949 and Q03164-PRO_0000390950 as first-class identifiers on gene-kmt2a and its fragment entries, so that any future query built from the graph's own accession fields inherits them?
Every source behind this entry (3)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    Fetched 2026-09-21, the IntAct REST endpoint ws/interaction/findInteractions returns 276 distinct interaction records for Q03164 against 142 for PSICQUIC id:Q03164, with containment in that direction and no record unique to PSICQUIC; the same comparison gives MEN1 133/58, PSIP1 113/49, DOT1L 60/36, AFF4 161/133 and AFF1 79/66, and the records visible only to REST include EBI-15969392 (KMT2A-MEN1, MI:0407 direct interaction, menin stored as O00255-2), EBI-2614630 (MLL-N Q03164-PRO_0000390949 with MLL-C Q03164-PRO_0000390950) and all six MEN1-JUND records.

    What this evidence cannot show

    The two endpoints do not return the same unit - PSICQUIC returns MITAB rows and REST returns binary-interaction documents - so the comparison is made on distinct IntAct interaction accessions (EBI-nnnnnn), which both expose, rather than on row counts. Four query forms were run against PSICQUIC as controls and are printed in the analysis output. One day, one database; IntAct also gains depositions, so the absolute counts will drift while the containment relation should not. The REST endpoint additionally returns mRNA-miRNA documents and fusion-construct interactors that no PSICQUIC protein query would match; those are excluded from the comparison, which counts only records where one side's base accession is the bait's.

    Evidence that argues against this (1)
    • PSICQUIC is a federated standard designed for cross-database querying, and its `id:` semantics are defined by the MIQL specification rather than by IntAct. Judged as an implementation of MIQL it may be behaving exactly as specified, in which case the defect belongs to the choice of instrument here and not to PSICQUIC. That distinction does not change the measurement, but it does change who should fix it. EMBL-EBI IntAct PSICQUIC 2026
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    This graph's cycle-34 experimental-role grading counted 290 partner pairs across the six core proteins - PSIP1 36, KMT2A 91, AFF1 35, MEN1 31, AFF4 71, DOT1L 26 - and reported that PSIP1 has exactly one bait-supported partner of thirty-six.

    What this evidence cannot show

    This is a citation to this project's own prior work, not an external source, and it is used only to establish what denominator the earlier analysis used - a fact readable from the committed file. It is NOT evidence that the earlier figure was wrong in its own terms: the grading of experimental role is sound and its conclusions about prey-only evidence stand on the records it saw. The url field carries no resolvable external target and the entry should be read as an internal cross-reference.

    Evidence that argues against this (1)
    • A larger denominator does not automatically make the older proportions wrong. If the records PSICQUIC misses are distributed the same way as the ones it sees, every proportion in those analyses is unbiased and only their confidence intervals are too narrow. Nothing here tests that, and the records it misses are visibly NOT a random subset - they are concentrated in processed chains and isoforms - so the direction of the bias is unknown rather than shown. EMBL-EBI IntAct 2026
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    PMID 15199122 (Yokoyama A, Wang Z, Wysocka J, et al., Cleary ML, Mol Cell Biol 2004) is the founding MLL/menin complex paper whose IntAct curation this project has already adjudicated, and its records are among those filed against KMT2A processed chains.

    PMID 15199122DOI
    What this evidence cannot show

    Title, eight-author list, journal, year and DOI taken from a live NCBI efetch XML record fetched 2026-09-21 with the anchored PubmedData/ArticleIdList/ArticleId path (brief rule 81). Cited here only to identify the paper behind specific IntAct records; the graph's substantive reading of it, including the UTP18/WDR5 substitution, is in ana-utp18-is-not-in-the-mll-complex-it-is-wdr5-miscurated and is not restated.

    Evidence that argues against this (1)
    • This paper's IntAct curation is already known to contain a twenty-one-year-old substitution of UTP18 for WDR5, so it is an odd choice of exemplar for anything. It is used here for the narrow fact of which accession its records are filed under, which is independent of the substitution - but a reader should treat any other inference from these particular records with the same suspicion the earlier adjudication earned. Oncofusion knowledge graph, internal entry 2026
Connected to