NUP98 — domain architecture and residue coordinates (with the 1817/920 numbering trap resolved)

This is a prediction, not a measurement. No one has seen this protein in this shape. The colours say how much the software trusts itself, and that is worth reading before anything else — the orange stretches are places the model is effectively saying this part has no fixed shape.

• The FG repeats retained in every NUP98 fusion are the same elements that wild-type nucleoporins use to form the nuclear pore's permeability barrier. Anything that disrupts FG-repeat condensation is likely to disrupt nucleocytoplasmic transport in every cell in the body.
• The transforming element is intrinsically disordered, which means there is no pocket, no fold and no structure-based design route. Every therapeutic hypothesis for this fusion has to go through the partner or through downstream chromatin machinery.
• The 17-residue offset above canonical residue 392 is a live hazard: a breakpoint imported from the fusion literature into a canonical-numbered viewer will sit 17 residues to the left of its true position, which is enough to move it across an FG-repeat boundary.

1. What is the NUP98::NSD1 breakpoint in canonical residue numbering? UniProt annotates sites for five other NUP98 fusions but not this one, so it is the single missing coordinate here and it is obtainable from the published fusion cDNA.
2. The FG repeats are required for transformation, but 'required' was shown by deleting the whole region. How much of the FG region is the minimal transforming unit, and does that minimum differ between partners as the phase-separation model would predict?
3. GLEBS is retained in the fusions and binds RAE1. Does the fusion actually sequester RAE1, and if so does that contribute anything, or is GLEBS a passenger that happens to sit between the two FG blocks?
4. What is in NUP98-NUP96 residues 881-1331 and 1624-1817? Between them they are nearly 700 residues with no annotation in any database and no structure, and they sit inside a bar this graph currently draws as a single named entity. If the polyprotein's second product has further folds, the entry is hiding them behind one label.
5. If a generic phase-separating IDR substitutes for the NUP98 FG repeats, what is the FG repeats' non-redundant contribution in a patient fusion? The partitioning result says composition; nobody has tested whether an FUS-IDR chimera partitions the same coactivators, which is the experiment that would separate 'generic condensate' from 'FG-specific interactome'.

1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

   UniProt canonical NUP98 (P52948-1) is the 1,817-residue NUP98-NUP96 precursor with FG repeats 1 at 1-156, a GLEBS/RAE1-interaction region at 157-213, FG repeats 2 at 214-480, a peptidase S59 autoproteolytic domain at 738-880 and an autolytic cleavage site at 880/881; translocation sites are annotated at 391/392 (NUP98-RAP1GDS1), 486/487 (NUP98-HOXA9) and 531/532 (NUP98-KDM5A, -PHF23, -CCDC28A).

   UniProtKB entry P52948 (NUP98_HUMAN): reviewed sequence, length and positional feature annotations · UniProt Consortium · UniProtKB/Swiss-Prot 2026
   ▸What this evidence cannot show

   No NUP98::NSD1 site is annotated, so the breakpoint for the fusion this node assumes is bracketed rather than pinned. UniProt lists six isoforms; verified lengths are 1817, 1726, 937, 920, 1800 and 1714 for P52948-1 through -6.

   ▸Evidence that argues against this (3)

   • The peptidase domain is not good to +/-8 as this node states: it is good to about +/-17 at the N-terminus. CATH-Gene3D and SUPERFAMILY both start it at 730 against UniProt's 738, and the 1KO6 construct starts at 695 - 43 residues before the UniProt boundary. The C-terminal end is firmer, 880 (UniProt, PROSITE) to 887 (Pfam, InterPro, CATH, 2Q5X construct), and 880/881 is independently fixed by the autolysis cleavage site. InterPro Consortium; EMBL-EBI 2026
   • The bar named 'NUP96 (C-terminal polyprotein partner)' at 881-1817 is not a domain and should not be read as one. It is the second cleavage product of the polyprotein. Within it, the only annotated domain is the NUP96 C-terminal domain at 1332-1623 (Pfam PF12110 / InterPro IPR021967), with a CATH match at 1315-1406; residues 881-1331 and 1624-1817 carry no annotation at all. PDB 1KO6 models chain B as only 881-942 of it. InterPro Consortium; EMBL-EBI 2026
   • The GLEBS boundary disagreement this node already flags is confirmed and can be given a third value: UniProt 157-213, the RAE1-bound constructs 4OWR 157-213 and 3MMY 158-213, but Pfam PF21240 only 170-208 and CATH-Gene3D G3DSA:1.10.10.2360 158-213. The two crystal structures side with UniProt, so Pfam's 170-208 is the outlier here rather than the correction. InterPro Consortium; EMBL-EBI 2026
2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

   Domain databases place the peptidase S59 domain at 738-885 (InterPro IPR007230), 730-887 (CATH-Gene3D and SUPERFAMILY) and 739-885 (Pfam PF04096), and the GLEBS motif at 170-208 (Pfam PF21240), narrower than the UniProt 157-213.

   InterPro/Pfam/CATH-Gene3D/SUPERFAMILY domain matches for UniProt P52948 · InterPro Consortium; EMBL-EBI · InterPro 2026
   ▸What this evidence cannot show

   Model-match extents. Pfam's narrow GLEBS match is the outlier against both UniProt and the crystallised construct extents, so it should not be preferred simply for being tighter.

   ▸Evidence that argues against this (1)

   • Both crystal structures of the GLEBS-RAE1 complex contradict Pfam's narrow call, in UniProt's direction. SIFTS maps 3MMY chain B to P52948 158-213 and 4OWR chain B to 157-213, so the crystallographers needed 56-57 residues, not Pfam's 39. The RCSB instance record for 3MMY chain B shows author numbering 158-213 with only residues 174-178 unobserved - and that five-residue hole sits INSIDE Pfam's 170-208 window, so the part Pfam calls the motif is precisely the part with a disordered gap, while the flanks Pfam excludes are fully ordered and RAE1-contacting. PDBe / SIFTS et al. 2026
3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

   The C-terminal domain of human Nup98 is an autoproteolytic domain with a novel fold defining a new class of autocatalytic proteases, and the previously suggested nucleoporin RNA-binding motif within it is unlikely to bind RNA.

   The three-dimensional structure of the autoproteolytic, nuclear pore-targeting domain of the human nucleoporin Nup98 · Hodel AE, Hodel MR, Griffis ER, et al. · Molecular Cell 2002
   PMID 12191480DOI
   ▸What this evidence cannot show

   Structure of an isolated C-terminal domain; PDB 1KO6 ('Crystal Structure of C-terminal Autoproteolytic Domain of Nucleoporin Nup98', title verified) models 695-880 plus the cleaved 881-941 tail.

   ▸Evidence that argues against this (1)

   • The 2002 model was built at 3.0 A and was superseded mechanistically six years later. The same domain was solved at 2.3 A for wild type and 1.9 A for the S864A precursor, with a specific autoproteolysis mechanism proposed via an N-O acyl shift, the driving force attributed to structural constraints at the scissile bond generated by anchoring a conserved phenylalanine into a hydrophobic pocket, and Tyr866, Gly867 and Leu868 identified as providing most of the domain-to-tail contacts. That paper's own framing is that Nup98 may represent a new SUBTYPE of protein using autoprocessing - a subtype claim, not the new-protease-class claim carried here. Sun Y et al. 2008
4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.review

   NUP98 translocations involve the intrinsically disordered N-terminal region of NUP98 with over thirty partner genes, and leukaemogenesis is mediated by changes in chromatin structure and gene expression. The FG repeats form condensates, but condensation alone is not the mechanism: the partner-derived chromatin reader supplies targeting while the FG repeats set which coactivators the condensate partitions.

   Mechanistic insights and potential therapeutic approaches for NUP98-rearranged hematologic malignancies · Michmerhuizen NL, Klco JM, Mullighan CG · Blood 2020
   PMID 32766874DOI
   ▸What this evidence cannot show

   A review. The phase-separation mechanism is presented as a plausible and partly supported model rather than as settled, and the partner dependence of cofactor recruitment is explicitly noted as unresolved. CORRECTION 2026-09-19: this item previously ended 'possibly via phase separation', which the counter-evidence below shows is not a sufficient account. In NUP98-PHF23 the PHD finger, not the FG repeats, targets the condensate to H3K4me3/2 genes, while the FG repeats determine which coactivators partition into it (PMID 39922194); and replacing the FG repeats with an unrelated phase-separating IDR from FUS reproduces the transcriptional effect (PMID 34163069), so the requirement is generic condensation plus specific partitioning, not phase separation per se.

   ▸Evidence that argues against this (1)

   • Phase separation by itself is not the mechanism; what the condensate selects for is. Dissection of NUP98-PHF23 shows the PHD finger - not the FG repeats - targets the condensate to H3K4me3/2-marked developmental genes, while the FG repeats determine condensate COMPOSITION, selectively partitioning KMT2/MLL H3K4 methyltransferases, histone acetyltransferases and BRD4. Condensation without the right partitioning does not activate the oncogenic programme, so 'possibly via phase separation' understates a two-part requirement - chromatin targeting plus selective coactivator partitioning - and overstates liquid-liquid phase separation as a sufficient account. Ahn JH et al. 2025
5. E5strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

   The 920-residue NUP98 isoform (P52948-4, RefSeq NP_624357) is identical to the 1,817-residue canonical precursor over residues 1-392 and then lacks canonical residues 393-409, so canonical numbering equals short-isoform numbering up to 392 and exceeds it by exactly 17 thereafter.

   Pairwise comparison of UniProt isoform sequences P52948-1 (1817 aa) and P52948-4 (920 aa), cross-checked against RefSeq NP_624357 (Nup98 isoform 2 precursor, 920 aa) · Computed by us from UniProt and NCBI RefSeq sequence records · UniProt / NCBI RefSeq 2026
   ▸What this evidence cannot show

   This resolves the offset between these two specific records. Individual papers may number against yet another NUP98 construct; the 937 aa RefSeq NP_005378 is a third distinct length. Always check the length a source is using before importing a residue number.

   ▸Evidence that argues against this (1)

   • The +17 rule is keyed to two records Ensembl does not privilege, and it is fragile in both directions. The live Ensembl gene record for NUP98 lists 54 protein-coding transcripts with 44 distinct protein lengths from 93 to 1848 aa; the Ensembl CANONICAL transcript, ENST00000324932, encodes 1800 aa across 33 exons - 17 residues SHORTER than the 1817-aa UniProt canonical, so the same 17-residue quantum reappears at a completely different comparison. In the 900-940 aa window alone Ensembl has ten transcripts at 903, 911 (x2), 912 (x2), 913, 919, 920 (x2) and 937 aa, so 'the 920-residue isoform' is not a unique address and a paper numbering against any of the other nine will not obey +17. Ensembl; EMBL-EBI 2026
6. E6strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

   NUP98-NSD1 induces AML in vivo, enforces HoxA7/A9/A10 and Meis1 expression, and requires its NUP98 FG-repeat domain: deleting the FG-repeat domain precluded Hox-A gene activation and myeloid progenitor immortalisation.

   NUP98-NSD1 links H3K36 methylation to Hox-A gene activation and leukaemogenesis · Wang GG, Cai L, Pasillas MP, Kamps MP · Nature Cell Biology 2007
   PMID 17589499DOI
   ▸What this evidence cannot show

   Mouse bone-marrow transduction and in vitro immortalisation assays; the FG-deletion control establishes necessity in that system, not the precise residue boundaries of the required element.

   ▸Evidence that argues against this (1)

   • Necessity in that assay does not mean the FG repeats are doing anything FG-specific. An artificial HOX chimera in which the NUP98 phenylalanine-and-glycine repeats were REPLACED by an unrelated phase-separating intrinsically disordered region from FUS had similar enhancing effects on genome-wide chimera binding and target-gene activation; the intrinsically disordered regions are needed to establish phase-separated puncta and to induce leukaemic transformation, and phase-separated NUP98-HOXA9 creates CTCF-independent chromatin loops enriched at proto-oncogenes. The FG deletion therefore removed a generic condensate-forming capacity, and its result cannot distinguish that from an FG-specific interaction. Ahn JH et al. 2021
7. E7established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

   An AlphaFold model AF-P52948-F1 covers all 1,817 residues.

   AlphaFold Protein Structure Database entry for P52948 (model AF-P52948-F1) · Jumper J, Varadi M, et al.; EMBL-EBI/DeepMind · AlphaFold DB 2026
   ▸What this evidence cannot show

   The entire FG-repeat region retained in fusions is intrinsically disordered, so the AlphaFold model of the transforming part of this protein carries essentially no structural information. Experimental structures exist only for the GLEBS-RAE1 complex (3MMY, 4OWR) and the peptidase domain (1KO6, 2Q5X).

   ▸Evidence that argues against this (1)

   • The model is among the least informative in this set of protein entries. Per-residue pLDDT recomputed from the v6 file gives 942 of 1,817 residues (52%) below 50 and only 90 residues (5%) at or above 90, with a mean pLDDT of 55.7 - so more than half the chain, including essentially all of the FG region retained in the fusions, sits in the band AlphaFold flags as probably disordered. 'Covers all 1,817 residues' and 'tells you anything about 1,817 residues' are different statements, and only the first is supported. AlphaFold DB / EMBL-EBI; PDBe SIFTS 2026
8. E8strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

   The autoproteolytic S59 domain's boundaries spread over seventeen residues across sources - CATH-Gene3D G3DSA:3.30.1610.10 and SUPERFAMILY SSF82215 both give 730-887, InterPro IPR007230 738-885, Pfam PF04096 739-885 and UniProt with PROSITE PS51434 738-880 - and the crystallised CONSTRUCTS are wider still in SIFTS: 1KO6 maps to P52948 695-880 and 2Q5X to 733-887 - but see the correction below, because 1KO6's observed coordinates start at 729, not 695. The NUP96 bar stored at 881-1817 is a cleavage product, not a domain: the only domain InterPro annotates in it is IPR021967 / Pfam PF12110 at 1332-1623.

   InterPro member-database domain matches for UniProt P52948 (Pfam, PROSITE, CATH-Gene3D, SUPERFAMILY, PANTHER) · InterPro Consortium; EMBL-EBI · InterPro 2026
   ▸What this evidence cannot show

   Profile-model extents and deposited construct ranges, not measurements; SIFTS reports the construct rather than the observed model. Retrieved 2026-09-19 (database records; no full text read). None of this changes a fate assignment, because the NUP98::NSD1 breakpoint cluster lies hundreds of residues upstream at 486-531. CORRECTION 2026-09-19: an earlier version of this item used the 1KO6 SIFTS construct range (695-880) as evidence that crystallised extents are WIDER than the profile models. The observed coordinates of 1KO6 chain A run 729-880 (auth 677-711 unobserved, UniProt = auth + 17), which is inside the profile range, so 1KO6 does not support that inference. 2Q5X (733-887 observed except 804-807) does.

   ▸Evidence that argues against this (1)

   • 1KO6 does not model 695-880, and the item's conclusion inverts once that is fixed. The RCSB entity alignment puts entity residue 2 at UniProt 695 and the auth mapping puts entity 1 at auth 677, so UniProt = auth + 17; the instance record then shows auth 677-711 unobserved, i.e. UniProt 694-728 has no coordinates. The real observed extent of 1KO6 chain A is 729-880 - 34 residues shorter at the N-terminus than the SIFTS construct mapping quoted here, and squarely INSIDE rather than outside the profile-model range it was cited to widen. 2Q5X survives better: auth 716-870 = UniProt 733-887 with only 804-807 unobserved, and at 1.9/2.3 A against 1KO6's 3.0 A it is the better structure to reason from. RCSB Protein Data Bank 2026

• refutes →NUP98::NSD1, t(5;11)(q35;p15.5)

  Counter-evidence added cycle 41: the FG-repeat requirement is not FG-specific. Replacing the NUP98 FG repeats with an unrelated phase-separating IDR from FUS gave similar enhancement of genome-wide chimera binding and target-gene activation (Ahn 2021, PMID 34163069), and condensation alone is insufficient - the PHD finger targets the condensate to H3K4me3/2 genes while the FG repeats set its composition (Ahn 2025, PMID 39922194).

• related to →NUP98::NSD1, t(5;11)(q35;p15.5)

  Counter-evidence (InterPro P52948): the peptidase domain's N-terminal tolerance is ~17 residues, not the +/-8 stated, and the bar labelled 'NUP96' at 881-1817 is a polyprotein cleavage product rather than a domain -- the only annotated fold inside it is 1332-1623. None of this moves the NUP98::NSD1 breakpoint bracket at 486-531.

• part of →NUP98::NSD1, t(5;11)(q35;p15.5)

  5-prime partner domain architecture; the FG-repeat region is retained and the peptidase domain lost

• part of →NUP98::KDM5A (NUP98-JARID1A), t(11;12)(p15;p13)

  same NUP98 FG-repeat contribution with a different 3-prime partner

• depends on →NUP98 fusions share the menin dependency (confirmed preclinically) — and the partner-supplied reader/writer module is not why

  supplies the residue coordinates the obligate-interface argument is drawn on

• supports →Mechanism class 5 — enhancer hijacking (inv(3) GATA2/EVI1 — and DUX4-r, which only half belongs here)

  the retained FG repeats are required for HOXA activation and myeloid immortalisation

• tests →NUP98 fusions share the menin dependency (confirmed preclinically) — and the partner-supplied reader/writer module is not why

  the FG-deletion control shows the NUP98 side is necessary, framing whether the partner interface is obligate

• supports →Breakpoint anatomy, and an audit of the prognosis claim

  three annotated NUP98 breakpoints (392, 487, 532) include progressively more of FG-repeat region 2

• ← analogous toEWSR1 — domain architecture and residue coordinates

  both contribute an intrinsically disordered, phase-separating low-complexity region as the transforming module

• ← part ofThe nuclear pore and NUP98 fusions — a fusion that takes a pore protein's disordered repeats off the pore and builds nuclear droplets with them

  the retained FG-repeat IDR is the compartment-defining module

• ← part ofWhy a nucleoporin fusion needs menin-MLL1: the second menin-dependent fusion family, and what it borrows rather than carries

  the retained FG-repeat region is the NUP98 contribution to the fusion

• ← part ofNUP98::KMT2A, inv(11)(p15q23) — the NUP98 fusion that contains KMT2A and is menin-inhibitor RESISTANT

  retains FG repeats 1+2 and GLEBS

• ← related toNUP98 rearrangements are 7.2% of paediatric AML and 1.6% of adult, with 35% five-year overall survival — and the partner changes everything

  85% of breakpoints in NUP98 exon 12 or 13

• ← part ofNUP98::HOXA9, t(7;11)(p15;p15) — the founding NUP98 fusion, and an incomplete oncogene

  retains FG repeats 1+2 and GLEBS; 19- or 37-repeat isoforms

• ← part ofNUP98::HOXD13, t(2;11)(q31;p15) — the myelodysplasia fusion, and a warning against targeting p53

  GLFG repeats fused to the HOXD13 homeodomain

• ← part ofNUP98::DDX10, inv(11)(p15q22) — the 'helicase' fusion whose helicase activity turns out not to be the point

  FG repeats fused to DDX10 across inv(11)(p15q22)

• ← part ofNUP98::TOP1, t(11;20)(p15;q11~q12) — the most precisely mapped NUP98 breakpoint, and the reciprocal that was wrong twice

  NUP98 exon 13 to TOP1 exon 8 in all ten reported cases

• ← related toDDX10 domain architecture (UniProt Q13206)

  the two halves of NUP98::DDX10

• ← related toHOXD13 domain architecture (UniProt P35453)

  the two halves of NUP98::HOXD13

• ← related toTOP1 domain architecture (UniProt P11387)

  the two halves of NUP98::TOP1

• ← related toNUP98 - the mobile nucleoporin that leaves the pore to put H3K4me3 on haematopoietic promoters, and the shared 5' donor of seven fusions in this graph

  residue coordinates, isoform-numbering offset and domain-database disagreements live there and are not restated here