NUMA1 / NuMA (Q14980) - a nuclear-matrix protein that becomes the spindle pole, and the 5' partner of a variant APL fusion whose coiled-coil is required for oncogenicity

Accession, entry name, length and synonyms were read from the JSON response of an exact-gene query that returned exactly one reviewed record, asserted in code before anything was written. The FUNCTION and localisation statements are curated summaries over roughly a dozen and several dozen primary citations respectively, none of which was read here - so 'nine compartments' is a union over experiments, cell types and cell-cycle stages, not a description of any one cell at any one time. UniProt records no DISEASE comment for this entry, so the fusion biology in E3 and E4 comes from the primary literature and not from the UniProt record.

• A 2,115-residue protein annotated for nine compartments, no domains, and a structural role in two different organelles is a poor object for any specific mechanistic claim, and that undercuts the reason it entered this graph. If NuMA is a scaffold present in most compartments at some point in the cell cycle, its recovery in an affinity purification is weak evidence of a functional relationship with any particular bait.

Abstract-level for the source paper plus the curated IntAct row fields, read from this project's own migrated instrument (analyses/interactome-union-migrated-to-rest). Three baits from one untargeted atlas is one observation reported three times, not three independent lines of evidence, and the read-queue threshold that admitted NUMA1 (two or more baits OR two or more publications) cannot distinguish those cases. No haematopoietic cell, no fusion-expressing cell, and nothing about which side of the KMT2A breakpoint the contact would map to - the fusion loses everything C-terminal of roughly residue 1444, so a KMT2A contact is not necessarily inherited by KMT2A::AFF1. OpenCell tags ENDOGENOUSLY rather than overexpressing, which is a genuine strength and is why this is weak rather than dismissed.

• The obvious dismissal is not available, and a future reader should not reach for it. NuMA is an abundant structural protein and the natural objection is that it is a non-specific affinity-purification background hit - but this graph TESTED that hypothesis in cycle 43 (hyp-atlas-only-spokes-are-predicted-by-prey-abundance-not-by-the-bait) and it was NOT supported in the form that matters: abundance moves low-confidence-spoke membership by an odds ratio of only 1.04 per standard deviation, and what predicts high-confidence membership is the prey:bait ratio. So NUMA1 cannot be waved away on abundance grounds; the honest weakness is evidence quality - one paper, one method, one non-haematopoietic host, sub-threshold score.
• Being recovered against three baits may argue AGAINST specificity rather than for it. KMT2A, MEN1 and PSIP1 are all nuclear and all in the same scaffold, so a shared partner is expected under either a real-complex model or a shared-background model, and nothing in these three rows discriminates. A partner recovered against one bait with two independent publications would be better evidence than this.

Abstract-level across five papers, all fetched and verified 2026-09-22: 9288109 (Wells, founding), 12584566 (Dong, the dimerization-domain requirement), 14737102 and 18408763 and 21255834 (Sukhai, transgenic models), 28449810 (Zou, NUMA1-PDGFRB). No full text or figure was read, so the magnitudes, the exact deletion boundaries and the controls are not established here. NuMA-RARA is a RARE variant: PML-RARA accounts for more than 99% of APL, so this is a small number of patients and the transgenic work is a model rather than the human disease. The relevance to KMT2A::AFF1 is an ANALOGY about fusion architecture - N-terminal partner supplying an oligomerisation domain - and not a shared mechanism, lineage or partner.

• The analogy may mislead in the direction this graph is already inclined to go. NuMA-RARA works by dimerising a nuclear RECEPTOR and recruiting a corepressor to retinoid response elements; KMT2A::AFF1 works by mislocalising an elongation machinery to HOXA chromatin. That both have an N-terminal coiled-coil-containing partner is a similarity of architecture, not of mechanism, and using it to argue that KMT2A's N-terminus 'must' matter by oligomerisation would be reasoning from shape rather than from evidence.
• The Wells 1997 reading is explicitly deflationary about the 5' partner's importance. The paper argues its results suggest that interference with retinoid signalling, NOT disruption of the partner's own organisation, is what is essential to the APL phenotype - which is the opposite of the 'the 5' partner sets the biology' inference this entry is otherwise assembled to support. Dong 2003's dimerization-domain requirement cuts the other way, and the two have not been reconciled.
• Recording NUMA1-PDGFRB alongside NuMA-RARA risks inflating the analogous-fusion dimension with a single case report. The PDGFRB fusion is one patient in one paper with no functional characterisation, and its imatinib response cannot establish that the fusion was the driver.

Abstract-level. This is ONE system - COS cells, a monkey kidney fibroblast line - with V5-epitope-tagged constructs expressed from transfected plasmid, so every address reported is an overexpressed tagged protein in a non-haematopoietic non-human cell. The contradiction with Wells 1997 and Sukhai 2011 is NOT resolved in the literature and is not resolved here; the most likely explanation is the system, but that is an inference. No DOI was retrievable for this record from the efetch response, so it is cited by PMID and title only.

• If the cytoplasmic result is a tagged-overexpression artefact, then so is the four-way comparison built on the same constructs, and the comparison is what makes this item valuable. The panel cannot be half-trusted: either the system reports addresses faithfully, in which case NuMA-RARalpha is cytoplasmic and Wells 1997 is wrong, or it does not, in which case the clean PML/PLZF/NPM/NuMA distinction is not evidence about the fusions either.
• This directly indicts several localisation records elsewhere in this graph. Entries here take subcellular and subnuclear addresses from epitope-tagged overexpression experiments without flagging the expression system as a confound, and this is a worked case where that choice produces a result contradicted by patient material and by an animal model. The same scepticism should be applied to this graph's own tagged-construct localisation claims.

ID-set comparison at retmax=600 is complete for every set here, the largest being 621, so nothing was truncated - which matters because a truncated union was computed and discarded earlier in this same cycle on a different gene. Title-and-abstract scope only. The NMP22 disjointness is a measurement of token overlap and not a proof that no NMP22 record concerns this protein; 'essentially disjoint' means zero overlap with NuMA and one with NUMA1, not zero relevance. 'NuMA' is a short token and may pick up unrelated uses, which would inflate the 621 - individual titles were not screened at that scale.

• The bare token NuMA is four characters and case-insensitive in PubMed, so the 621 is an upper bound that may include acronym collisions this measurement did not screen. If a meaningful fraction is noise then the alias split is less severe than stated - though the direction would not change, since NUMA1 and NuMA share only 15 records either way.