Do NUP98's seven fusions resemble each other more than they resemble their partners' other fusions? The test cannot be run for six of them, and the seventh resembles neither side
also called the NUP98 shared-address test · cross-partner scan
NUP98 is stuck onto seven different partner genes in seven different leukaemia fusions, which makes it the best case for asking whether a fusion family is really about the shared piece. The answer turned out to be that the question cannot be asked properly here - six of the seven partners appear in no other fusion this project has written - and the one partner that can be compared behaves unlike both groups.
The same thing, for a scientist
**A pre-specified test, run, and returning a negative.** `concept-low-grade-ess-is-a-nua4-to-polycomb-fusion-family` argues that a fusion family can be defined by **the complexes its halves belong to** rather than by a shared gene, and that the analogue of a KMT2A fusion might be **any address-to-effector fusion**. NUP98 is the sharpest available test - one address donated to seven effectors in this corpus - and the prediction is that the seven should resemble each other in mechanism **more** than each resembles its own 3' partner's other fusions. **Result 1. The comparison set is empty for six of seven.** Across all 52 fusion entries in this corpus, only three genes appear as a 3' partner in more than one entry (KMT2A 4, MLLT10 2, NUTM1 2). Of NUP98's seven partners, **only KMT2A has other fusions here** - `fusion-ldb1-kmt2a`, `fusion-vim-kmt2a`, `fusion-yap1-kmt2a-yap1`. DDX10, HOXA9, HOXD13, KDM5A, NSD1 and TOP1 have none [E1]. **Result 2. The one runnable case is an outlier on both sides.** `fusion-nup98-kmt2a` records blasts with **very low** Hoxa5/a7/a9/a10, Meis1, Eya1 and Pbx3 against normal marrow or KMT2A-ENL/KMT2A-AF9 blasts, and resistance to MI-2-2 at 3-12 uM and to JQ1 - because the inversion deletes KMT2A exon 1 and with it the menin-binding motif [E2]. It looks like neither the NUP98 family (HOXA/MEIS1-high, menin-sensitive) nor the KMT2A family (menin-dependent). **Result 3. The axis on which all seven cohere is non-discriminating by construction.** A keyword scan finds condensate/FG-repeat language in **7 of 7** NUP98 fusion entries and **0 of 4** non-NUP98 comparators (the fifth at 2), with no length confound [E1]. But the FG module **is** NUP98, so a plain named-gene family predicts that split exactly as strongly as a complex-defined family does. **The one axis that separates the sets perfectly is the one on which the two competing framings agree** [E1]. **Result 4, and it is the one that actually bears on the claim.** This corpus's own `mechanism-nup98-fusion-menin-mll1-dependency` records that NUP98-fusion menin sensitivity is **partner-graded** - about a six-fold IC50 spread, one of three PDX refractory, NUP98::KMT2A resistant outright - and that the load-bearing downstream node is MEIS1 rather than HOXA, because NUP98-HOXD13 cells lose no Hoxa occupancy at all [E3]. **A dependency that varies across the seven is evidence against the strong form of 'the seven resemble each other most.'** **Verdict: negative, and reported as one.** Q123's framing is not refuted - nothing here touches the ESS biochemistry it rests on - but NUP98 does not supply the support this beat hoped to find, and the sharpest available test came back empty.
Why the empty comparison set is itself the most useful output
The brief expected a scoreable comparison. What the scan found first is that this corpus is organised around 5' donors and almost never sees a 3' partner twice: 52 fusion entries, and exactly three genes recur as a 3' partner 1.
That is a standing limit on the PI's fifth dimension, not a defect of this beat. 'The analogous biology in other fusion families' has been pursued here by comparing named-gene families - the KMT2A family, the NUP98 family, the ETV6 family - and those are all 5'-donor families. A comparison that needs a shared 3' partner has almost no material to work with, and nobody had noticed because nobody had counted.
What this does NOT show. It is a statement about this corpus, not about the literature. DDX10, HOXA9, HOXD13, KDM5A, NSD1 and TOP1 may each appear in other fusions that this graph has not written; no PubMed query was run to bound that and none is claimed. Establishing which of the six are genuinely partner-unique and which are merely unwritten is a cheap, well-defined next beat, and it is the prerequisite for ever running this test properly.
The instrument, and the one place it would have inverted the answer
The scan counts regex hits over the whole serialised JSON of each entry on four axes - HOXA programme, menin dependence, condensate/FG, DOT1L/H3K79 - and the full table is in analyses/q126-the-partners-named-but-never-written/nup98-shared-address-cross-partner-scan.json.
It measures what this corpus's beats wrote, not biology, so only the categorical zero/non-zero pattern is used and the magnitudes are discarded. The length control is stated because it is the obvious confound and it does not fire: the non-NUP98 comparators run 17,469-37,231 characters against the NUP98 entries' 15,242-49,672, so their zeros are not short-entry artefacts 1.
And the specific failure this instrument would have produced if trusted. fusion-nup98-kmt2a scores 36 on the menin axis - the second-highest of the seven - and the content of the entry is that the fusion is menin-inhibitor RESISTANT 2. A count of the string 'menin' cannot see the sign of a claim, and on the single most informative entry in the set it would have inverted it. That is why Result 2 was read from the entry's prose and not taken from the table, and it is rule 92's instruction applied before rather than after the verdict.
| script | analyses/q126-the-partners-named-but-never-written/nup98-shared-address-cross-partner-scan.json |
|---|---|
| status | run |
| result summary | NEGATIVE on the proposed test. Six of seven NUP98 partners have an empty comparison set in this corpus; the seventh (KMT2A) resembles neither family; the only axis on which all seven cohere is the FG/condensate module, which both competing framings predict equally and which therefore carries no discriminating information; and the corpus's own best-evidenced cross-partner mechanism, menin-MLL1 dependency, is explicitly partner-graded, which cuts against the strong form of the claim. |
| pre specified | The prediction was written in the dispatch brief before the scan was run, together with the instruction not to force it. The scan was run once and its first result - the empty comparison set - was not the result the prediction anticipated and is reported first. |
| criterion informativeness clause | Cycle 86's rule asks what result would make a criterion itself uninformative. For this test it is: IF the axis that separates NUP98 fusions from non-NUP98 fusions is an axis the shared GENE supplies, then a clean separation confirms only that NUP98 fusions contain NUP98, and the criterion is uninformative. THAT IS WHAT HAPPENED - the perfect 7/7-versus-0/4 split is on the FG/condensate axis - so the separation is reported as non-discriminating rather than as a pass. |
inputs
[ "all 52 fusion entries in data/graph/nodes", "fusion-nup98-kmt2a", "mechanism-nup98-fusion-menin-mll1-dependency", "concept-low-grade-ess-is-a-nua4-to-polycomb-fusion-family" ]
what would make it runnable
[ "Establish, by literature search, which of DDX10, HOXA9, HOXD13, KDM5A, NSD1 and TOP1 genuinely have no other fusion partner and which are merely unwritten here. This was NOT done and is the prerequisite.", "Replace keyword counts with a fixed mechanism-feature schema in each fusion entry's data block.", "Use the per-partner menin IC50 spread as the quantitative axis; it exists and was not pulled in this beat." ]
corpus snapshot
{
"node_files": 2374,
"fusion_entries": 52,
"date": "2026-09-23"
}- The most likely misuse of this entry is to read 'NUP98 fusions all have condensates and other fusions do not' as a therapeutic rationale for an FG-repeat- or condensate-directed agent. The 7-of-7 versus 0-of-4 split is a fact about which gene each fusion contains, not a demonstration that condensation is the shared vulnerability, and gene-nup98 records that the same FG region carries wild-type NUP98's own chromatin functions in normal haematopoietic progenitors.
- The second likely misuse is to read the NUP98::KMT2A outlier as licensing a general claim that menin inhibition fails in NUP98 disease. It does not: the corpus records menin sensitivity across the rest of the family and eviction of the fusion protein itself. NUP98::KMT2A is one fusion, two patients, one mouse and a micromolar tool compound.
- And the honest structural risk: an analysis whose headline is 'the test cannot be run' can be used to justify not running the follow-up. The follow-up here is cheap and specific - a literature search for other fusions of six named genes - and the value of this entry is entirely contingent on somebody doing it.
- Which of DDX10, HOXA9, HOXD13, KDM5A, NSD1 and TOP1 actually has no other fusion partner in the published literature? This scan establishes only that this corpus has none. The answer decides whether the empty comparison set is a fact about fusion biology - 3' partners genuinely do not recur - or a curation gap, and those are opposite conclusions drawn from the same table.
- Is 3'-partner non-recurrence a real property of oncogenic fusions? If a 5' address is promiscuous and a 3' effector is not, that is a structural claim about how fusion oncogenes are built, and it would be independently interesting. Nothing in this beat tests it; the observation that only three genes recur as 3' partners across 52 entries is the prompt, not the answer.
- Would the per-partner menin IC50 spread, scored properly, discriminate the two framings? A six-fold spread across the NUP98 family is a quantitative comparator this corpus already holds. If the spread within the NUP98 family is smaller than the spread between families, the shared-address framing gains support; if not, it does not. This beat did not pull the numbers and states no expectation.
- Does the empty-comparison-set result generalise to ETV6 and TCF3, the two other multi-fusion 5' donors here? Both have exactly two entries, so the same test would be even weaker for them - but running it would establish whether NUP98 is unusually well covered or unusually badly compared.
runs that concept's transferable claim - that the analogue of a KMT2A fusion may be any address-to-effector fusion - against NUP98's seven fusions. Returns a negative on the test's feasibility, not a refutation of the concept
- related to →NUP98::KMT2A, inv(11)(p15q23) — the NUP98 fusion that contains KMT2A and is menin-inhibitor RESISTANT
the only NUP98 fusion whose 3' partner has other fusions in this corpus, and it resembles neither family
the axis that would discriminate the two framings; it is partner-graded, which cuts against the strong form of the shared-address claim
the shared 5' donor whose seven fusions the scan compares