Oncofusion

start / graph / ana-intact-holds-forty-seven-fusion-side-records-our-instrument-drops-and-every-partner-is-already-written

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

IntAct holds 51 interaction records whose participant IS a KMT2A fusion across 6 constructs, our coverage instrument dropped every one of them silently, and closing the single gap they contained took the partner list to 21 of 21

also called fusion-construct interactors · IntAct fusion-side records · EBI-2890930 mll_af9

This project measures how much of the fusion protein's partner list it has written up, and the program doing the measuring was silently skipping the very records where the partner was measured against the fusion itself. Reading them found fifty such records, five different fusion constructs, and exactly one protein the graph had never written down.

The same thing, for a scientist

**UPDATED CYCLE 55, and the update is the point: cycle 54 said 47 rows and 18 partners were a FLOOR, and they were.** Making the parser's silent skip LOUD surfaced a **fifth** fusion construct, `human_mll_gas7_fusion_protein` (EBI-1157918), taking the count to **50 rows and 21 distinct partners, 20 of which have an entry**. The one that does not is **SFPQ** (P23246), an anti-tag co-IP partner of the MLL-GAS7 construct in HEK293 from PMID 26766589 — a *Cancer Cell* 2016 paper this graph already cites **ten** times, including in `gene-prmt1` and `gene-kdm4c`. So the biology was known and the fusion-side rows were not. And the skip was dropping **three** classes, not one: fusion constructs, Ensembl **gene**-record participants (hoxa9, hoxa7, meis1, myc, msx2 — regulatory relationships curated as interactions) and **microRNAs**. 116 row-sides across 39 forms. The coverage figure is unchanged at 94/217 = 43.3%, deliberately: resolving these into the numerator would change what the number means mid-series. **The instrument gap is real, and the one content gap is now closed:** gene-sfpq was written the same cycle and a re-run reports **21 of 21 partners with an entry, 0 without**, per-construct counts unchanged. Writing it turned up the thing the enumeration could not see -- SFPQ is itself the 5' partner of **SFPQ::ABL1** in Ph-like B-ALL, which localises to the **nucleus** and runs a different signalling network from BCR-ABL1 with the same kinase (PMID 35061886).

Facts and figures
scriptanalyses/intact-fusion-construct-interactors/run.py
resultsanalyses/intact-fusion-construct-interactors/results.json
rows kept51
rows dropped by the substring guard4
distinct partners21
partners with an entry21
partners without an entry0
the partner listAFF1, AFF4, CBX8, CCNT1, CDC73, CDK9, CTR9, DOT1L, HOXA9, KHDRBS1, LEO1, MED19, MED26, MEIS1, MLLT1, PAF1, PAX5, PRMT1, RNF2, SFPQ, SKIC8
why the instrument missed themA fusion-construct interactor has an empty `interactorXrefs`, so core-complex-interactome-gap's gene_of() regex, which extracts a symbol from a '(gene name)' suffix, returns None and the row is skipped without a warning. The fix is not to change the coverage figure but to make the skip LOUD.
the normalisation trap this script fell intoFirst run reported HOXA9 and MEIS1 as absent from the graph because IntAct names them hoxa9_mouse_gene and meis1_mouse_gene and the normaliser left HOXA9MOUSEGENE. gene-hoxa9 and gene-meis1 both exist. Corrected before the result was written up; the count went 16/18 to 18/18.
what is NOT establishedA curated interaction with a transfected tagged construct is evidence about the construct, not automatically about the patient lesion, and several of these rows are mouse. Species is kept per row in results.json and must be read before any human claim.
shown for fusiontrue
absence classmeasured-not-analysed
rows dropped as duplicate experiments3
n distinct constructs6
n construct acs7
figure provenanceCycle 54 measured 47 rows / 18 partners over four hand-found constructs. Cycle 55 made the parser's silent skip loud (50 rows, 21 partners, 1 gap), wrote the missing entry, then ran a systematic census that found a sixth construct and a duplicate record: 51 rows, 6 constructs, 21 partners, 21 with an entry, 0 without. All three generations of the figure are kept here deliberately, because the movement is the finding.
constructs
{
  "EBI-2890930": "mll_af9_human_protein, 32 interactions",
  "EBI-2620068": "mll_enl_human_protein, 10",
  "EBI-2620048": "mll_aff4_fusion, 7",
  "EBI-2620075": "mll_aff4_fusion-1, 2",
  "EBI-1157918": "human_mll_gas7_fusion_protein, 3",
  "EBI-9513644": "mll_ell_fusion_protein, 1 -- the sixth, found by the census",
  "EBI-11579188": "human_mll_gas7_fusion_protein AGAIN -- a duplicate interactor record for EBI-1157918; its 3 rows share interaction ac EBI-11579418 with the original and are dropped rather than counted"
}
extended cycle 55 the floor was real
{
  "what_happened": "Cycle 54 named four construct ids found by hand and said explicitly that 47 rows and 18 partners were a FLOOR. Cycle 55 taught core-complex-interactome-gap to REPORT the participants its parser drops instead of skipping them silently, and the report immediately proved the floor.",
  "a_fifth_construct": "`human_mll_gas7_fusion_protein`, EBI-1157918, which cycle 54 never reached. The fourth record EBI-2620075 is also now identified as `mll_aff4_fusion-1` rather than an unknown.",
  "new_totals": "50 rows, 21 distinct partners, 20 with an entry, 1 without.",
  "the_one_gap": "**SFPQ** (P23246). Reported to associate with the MLL-GAS7 fusion by anti-tag co-IP in HEK293, human, from PMID 26766589 -- and it has no entry. Its two co-partners on the same construct and the same paper, KHDRBS1 (Q07666) and PRMT1 (Q99873), both do. -- CLOSED the same cycle: gene-sfpq now exists and the re-run reports 21 of 21 partners with an entry, 0 without. The entry's centre of gravity turned out not to be the MLL-GAS7 row at all (typed `association`, the weaker term, in HEK293, from a paper whose abstract never names SFPQ) but SFPQ::ABL1, a Ph-like BCP-ALL fusion in which SFPQ is the 5' partner.",
  "the_source_is_already_well_cited": "PMID 26766589 is Cheung N, Fung TK, Zeisig BB, Holmes KB et al., 'Targeting Aberrant Epigenetic Networks Mediated by PRMT1 and KDM4C in Acute Myeloid Leukemia', Cancer Cell 2016, doi 10.1016/j.ccell.2015.12.007 -- PRMT1 an H4R3 methyltransferase, necessary but NOT sufficient for leukaemic transformation, requiring co-recruitment of the H3K9 demethylase KDM4C by chimeric transcription factors. This graph already cites it in 10 entries including gene-prmt1 and gene-kdm4c. So the BIOLOGY was known; what was invisible were the fusion-side co-IP rows and the SFPQ partner in them.",
  "THREE_classes_were_being_dropped_not_one": "116 unresolved row-sides across 39 distinct forms, and they are not all fusion constructs: (1) fusion constructs (mll_af9 36, mll_aff4 13, mll_enl 10, mll_aff4-1 3, mll_gas7 3); (2) ENSEMBL GENE-RECORD participants -- hoxa9_human_gene ENSG00000078399 8, hoxa7_human_gene 6, meis1_human_gene 3, myc_human_gene 2, msx2_human_gene 1 -- which are regulatory or protein-DNA relationships curated as interactions, a different kind of claim; and (3) microRNAs, hsa-mir-17-5p and hsa-mir-92a-3p. The interactome coverage measure has never seen any of the three.",
  "the_coverage_figure_did_not_move": "94/217 = 43.3%, unchanged. The fix deliberately does NOT resolve these into the numerator, because that would change what the number means mid-series. It makes the drop loud so the next invisible class is noticed the first time.",
  "re_run_after_closing_the_gap": "50 rows, 21 partners, 21 with an entry, 0 without; per-construct totalElements 32 / 10 / 7 / 2 / 3, identical to the first run, so the only thing that moved is the graph."
}
How this could bite us back
  • This entry reports that an instrument silently skipped a class of record, and the honest finding is that skipping it cost nothing in coverage. It should not be cited as evidence that this project's interactome numbers are unreliable, nor as evidence that they are fine: it shows one specific blind spot whose contents happened to be already covered, which is luck plus prior work rather than a property of the instrument.
  • Nothing here is a therapeutic claim. It is a completeness audit of a database query.
Still open
  1. ANSWERED, cycle 55, and it paid for itself. gene-sfpq is written and the enumeration re-run now reports 21 partners, 21 with an entry, 0 without -- with every construct's totalElements identical to the earlier run, so the closure is a real change and not IntAct drift. What made it worth more than closing a gap: SFPQ is ITSELF the 5' partner of a fusion, SFPQ::ABL1 in Ph-like B-cell precursor ALL, exon 9 of SFPQ to exon 4 of ABL1 in all six reported cases (PMID 32306475), and the ABL1 intron-4 break leaves the chimaera without the ABL1 SH2 domain (PMID 21435002). PMID 35061886 then compares it directly with BCR-ABL1 -- same kinase, different 5' partner -- and finds SFPQ-ABL1 in the NUCLEAR compartment, a weaker proliferative driver, no PI3K/AKT activation, up-regulated cell-cycle/DNA-replication/spliceosome and down-regulated ErbB/NF-kB/VEGF/MAPK phosphosignalling, and Mcl-1/Bcl-xL dependence. That is this project's 5'-partner thesis measured on a kinase fusion by someone else. None of it was in the dispatch; it was found by asking whether the partner being written was itself a fusion partner.
  2. Should the Ensembl gene-record participants (hoxa9, hoxa7, meis1, myc, msx2) count as interactome coverage at all? They are regulatory relationships curated in an interaction database, and treating them as protein-protein partners would inflate the figure while excluding them silently -- as happens now -- hides that IntAct holds them.
  3. Should core-complex-interactome-gap's gene_of() be taught to keep fusion-construct interactors, or merely to WARN when it drops a row? Keeping them would put fusion-side evidence in the coverage figure for the first time; warning is the smaller change and prevents the next silent skip.
  4. PARTIALLY ANSWERED, cycle 55, and the answer is honest about its own limit. analyses/intact-fusion-construct-census/run.py asked IntAct's findInteractor endpoint instead of accumulating ids by hand: 22 queries, 5,009 interactors, 208 candidates matching the construct signature, 0 query failures. It found a SIXTH construct, mll_ell_fusion_protein EBI-9513644 (1 row, partner PAX5), and a DUPLICATE record, EBI-11579188, whose three rows share interaction ac EBI-11579418 with EBI-1157918 and would have inflated the count 50 -> 53. But it recovered only 4 of the 5 constructs already known - missing EBI-1157918, the very record whose duplicate it returned - so its recall is below 100% and it CANNOT support 'there is no seventh'. The floor is one construct lower and still a floor. Remaining routes: a full crawl of interactors with empty cross-references, or the PSI-MI dumps.
  5. ANSWERED, cycle 55: gene-pax5 E4 now records it. PAX5 is curated against mll_af9 and mll_enl by anti-tag co-IP in HEK293T, PMID 24735925 / IntAct EBI-9512250, and is 1 of only 7 of the 21 partners appearing on more than one construct -- the other six being AFF1, CDC73, CTR9, DOT1L, LEO1 and PAF1, all transcription-elongation machinery. Two things the enumeration did not show came out of writing it up: both PAX5 rows carry the SAME IntAct interaction id, so they are one experiment counted twice rather than two observations; and PMID 24735925 was already cited once in this graph, for a lineage-split proteasome mechanism, with its interaction content never read.
Every source behind this entry (1)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    IntAct's findInteractions endpoint returns 32 interactions for EBI-2890930 (mll_af9_human_protein), 10 for EBI-2620068 (mll_enl_human_protein), 7 for EBI-2620048 (mll_aff4_fusion) and 2 for EBI-2620075 — 47 rows after 4 are dropped because neither participant is the queried interactor. They resolve to 18 distinct partners (AFF1, AFF4, CBX8, CCNT1, CDC73, CDK9, CTR9, DOT1L, HOXA9, LEO1, MED19, MED26, MEIS1, MLLT1, PAF1, PAX5, RNF2, SKIC8) and all 18 already have a per-protein entry in this graph by the same gene-/protein- naming rule the core-complex coverage analysis uses.

    What this evidence cannot show

    The construct list is four ids found by a research agent and named explicitly here rather than discovered, because a free-text IntAct search for `mll` returns wild-type KMT2A and everything citing it — so 47 rows and 18 partners are a FLOOR, not a census, and further fusion-construct interactors may exist that this does not reach. Partner symbols are resolved by stripping IntAct's species and record suffixes, which collapses a mouse gene and its human orthologue into one row; per-row species is preserved in results.json and several of these constructs were characterised in mouse. The in-graph test counts the EXISTENCE of an entry, not its depth, and it counts an entry that merely names the protein as covered if the id matches. Endpoint counts will drift as IntAct is recurated.

    Evidence that argues against this (2)
    • The headline could be read as 'our coverage instrument is broken', and the enumeration says the opposite about coverage: 18 of 18 partners are written, so the figure the instrument reports was not wrong about the biology, only about its own completeness. Anyone quoting this entry as evidence that the interactome measure understates the graph should note the measure would not have MOVED — every one of these partners is already in the numerator via its wild-type record.
    • These are transfected tagged constructs, several assayed in mouse cells, so a partner recovered here is not thereby a partner of the patient lesion. The same objection this graph applies to the Skucha AP-MS network applies with more force: at least that network used isogenic human lines with a negative control, whereas the construct records are a heterogeneous set of individual co-immunoprecipitations across laboratories and species.
Connected to