The standing priority asks for localization on every interactor; the graph has it on 14% of genes and 24% of its most central ones, and DOT1L has none at all
also called PI dimension coverage · localization gap · cell type cell state coverage · subnuclear localization audit
The lab asks that every protein partner record where in the cell it sits and which cells it matters in, and about one gene entry in seven has any location at all - counting generously. The protein targeted by this field's second drug class has no location recorded anywhere in its entry.
The same thing, for a scientist
Measured across **306 gene entries**, counting a gene as covered if it has a localization field **or** an edge to a `localization-*` entry: | | n | % | |---|---:|---:| | localization field | 37 | 12.1% | | edge to a `localization-*` entry | 10 | 3.3% | | **either** | **44** | **14.4%** | | edge to a cell-type or cell-state entry | 33 | 10.8% | | `subnuclear_localization` field | 7 | 2.3% | | `subcellular_localization` field | 4 | 1.3% | | `cell_type` / `cell_state` field | 1 / 1 | 0.3% | **And it is not better where it matters more.** By edge degree: top 25 genes **6/25 (24%)**, top 50 16%, top 100 17%, bottom 150 11.3%. **The misses are not obscure genes.** `gene-dot1l` (degree 85) has **zero** localization content of any kind; `gene-psip1` (LEDGF, degree 60) has two incidental uses of 'chromatin-bound' and no record; `gene-wdr5` (degree 39) mentions nucleolar stress as a drug phenotype and never its own localization. **All three are well characterised in the literature**, so this is unwritten work, not an unknown.
The vocabulary exists and is unconnected
The graph holds 31 localization-* entries and 46 cell-type/cell-state entries. Some are excellent - the subnuclear beat produced entries on nuclear speckles, the lamina, chromocentres, the nuclear matrix and AF4 bodies, several of which honestly record that nothing is published for a KMT2A fusion.
But only 10 gene entries edge to any of them. The project built the categories and then did not assign genes to them. That is a different and much cheaper problem than missing research: for most of these genes the answer is in UniProt's subcellular location section and in the primary papers already cited elsewhere in this graph.
Why it was easy to miss. Every individual entry looks complete - it has evidence, caveats, counter-evidence, citations. The gap is only visible across entries, as a field that is almost never filled, and no checker in this repo looks for it. The validator enforces schema and links; nothing enforces the PI's dimensions.
What should be done, and what should not
Should: a beat that fills localization and, where known, subnuclear_localization for the top ~50 genes by degree, each with a resolving citation - UniProt SUBCELLULAR LOCATION with its accession, or a primary paper that actually imaged or fractionated it. That is roughly 42 genes of unwritten work on the most-used entries in the graph, and it directly serves the standing instruction.
Should NOT: an automated pass that copies UniProt's subcellular-location string into every gene entry. Three reasons, all of which this project has already been bitten by. UniProt localization is frequently annotated by similarity rather than measured. It is a statement about the wild-type protein, and this graph's subject is a fusion - the localization that matters for MLLT1 or AFF1 is where the oncoprotein goes, which is often not where the partner normally lives. And subnuclear localization is almost never in UniProt at all; it needs the imaging paper.
So the field to fill is not one field. localization can often be sourced quickly; subnuclear_localization requires reading; and whether it was shown for the fusion or only for the wild-type partner has to be recorded either way - this graph already has a shown_for_fusion field on 45 entries and it is the right precedent.
| generated | 2026-09-20 |
|---|---|
| question | The standing PI priority asks for every interactor to carry cell type, cell state, subcellular and subnuclear localization. How much of that exists? |
| method | Structured fields on gene entries, PLUS edges to dedicated localization-* and cell-type/cell-state entries, so a gene whose localization lives in a separate node counts as covered. Prose spot-checked on the three highest-degree misses to confirm the absence is real and not a field-versus-prose artefact. |
| gene entries | 306 |
coverage — the numbers behind the chart
{
"localization_field": {
"n": 37,
"pct": 12.1
},
"edge_to_localization_entry": {
"n": 10,
"pct": 3.3
},
"either": {
"n": 44,
"pct": 14.4
},
"edge_to_cell_type_or_state": {
"n": 33,
"pct": 10.8
},
"subnuclear_localization_field": {
"n": 7,
"pct": 2.3
},
"subcellular_localization_field": {
"n": 4,
"pct": 1.3
},
"cell_type_field": {
"n": 1
},
"cell_state_field": {
"n": 1
}
}by centrality
{
"top_25": [
6,
25,
24
],
"top_50": [
8,
50,
16
],
"top_100": [
17,
100,
17
],
"bottom_150": [
17,
150,
11.3
]
}vocabulary that exists
{
"localization_nodes": 31,
"cell_type_or_state_nodes": 46
}spot checks
{
"gene-dot1l": "degree 85. ZERO localization mentions anywhere in the entry. No field.",
"gene-psip1": "degree 60. Two mentions, both incidental uses of 'chromatin-bound' about the menin-KMT2A-LEDGF clamp, not a localization record. No field.",
"gene-wdr5": "degree 39. Four mentions, all about NUCLEOLAR STRESS as a drug phenotype, not about WDR5's own localization. No field."
}caveats
[ "Counting a gene as covered if it has a field OR an edge to a localization entry is deliberately generous; the true per-gene figure for a properly recorded, cited subcellular AND subnuclear localization is lower than 14.4%.", "Degree is a proxy for importance and a crude one - it partly measures how many beats have touched a gene rather than how central it is biologically.", "This measures THIS GRAPH, not the literature. A missing localization is not a claim that the localization is unknown; for DOT1L, LEDGF and WDR5 it is very well characterised, which is what makes the gap a gap rather than a limit." ]
- Do not fix this with an automated UniProt copy. UniProt localization is often annotated by similarity rather than measured, it describes the WILD-TYPE protein while this graph's subject is a fusion, and it almost never contains subnuclear detail. An automated pass would fill the field and destroy the meaning - and this project has made that class of mistake twice already with symbols and accessions.
- And do not read 14.4% as 14.4% of the biology being wrong. Every one of these entries is individually sound; what is missing is a dimension the standing instruction asks for and no checker enforces. The gap is invisible per-entry and only exists in aggregate.
- How many of the top-50 genes have a subnuclear localization that is actually published for the FUSION rather than the wild-type partner? That distinction is the difference between a quick fill and a real beat.
- Should the validator warn when a gene entry has no localization, as it warns on unlinked nodes? A warning rather than an error would make the gap visible every cycle.
- Do the 31 localization entries cover the compartments that matter, or were they written around whichever compartments had papers?
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis
Across 306 gene entries in this knowledge graph, 44 (14.4%) carry any subcellular localization information, counting both a data field (37) and an edge to a dedicated localization entry (10); subnuclear localization is recorded on 7 (2.3%), and 33 (10.8%) connect to a cell-type or cell-state entry, against a standing instruction that every interactor should carry all four dimensions.
pi-dimension-coverage: structured-field and edge audit of cell type, cell state, subcellular and subnuclear localization across gene entries · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Run 2026-09-20 over 306 gene entries and 7,444 edges. **The counting rule is deliberately generous** - a gene counts as covered on a bare field or a single edge, so the figure for a properly cited subcellular AND subnuclear record is lower than 14.4%. **Established for what this graph contains and for nothing else**: a missing localization is not a claim that the localization is unknown. Degree is a crude proxy for importance, partly measuring how many beats have touched a gene.
Evidence that argues against this (1)
- A low field-fill rate is not automatically a defect under this project's own 'one entry per thing' rule - localization could legitimately live in the 31 dedicated localization entries rather than being duplicated onto every gene. That defence is why edges were counted too, and it fails on the numbers: only 10 gene entries edge to those nodes, so the information is not living there either. Shi J et al. 2013
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis
The absence is not an artefact of localization being recorded in prose rather than in a field: of the three highest-degree gene entries lacking a localization field, gene-dot1l contains no localization statement of any kind, gene-psip1 contains only two incidental uses of the phrase chromatin-bound describing a complex rather than a location, and gene-wdr5's four matches all describe nucleolar stress as a drug phenotype rather than WDR5's own localization.
pi-dimension-coverage spot checks: prose scan of the highest-degree gene entries lacking a localization field · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Three entries were checked, not all 262 lacking a field, so this establishes that the gap is real at the top of the list rather than quantifying how often prose substitutes for a field lower down. The scan used a fixed pattern of localization words and would miss a localization stated in unusual wording - which, given this project's history with over-flagging regexes, is the error direction that would make the gap look worse than it is.
Evidence that argues against this (1)
- DOT1L's localization being unrecorded is less damning than it first sounds: DOT1L is a chromatin-acting methyltransferase and its nuclear localization is so assumed that no beat would think to state it. The genuinely missing item for this project is not 'nucleus' but WHERE IN THE NUCLEUS, and that is unpublished for most of these proteins in a KMT2A-fusion context - which the subnuclear beat's own entries already record as bounded absences. Shi J et al. 2013
HIGHEST-DEGREE GENE WITH NO LOCALIZATION AT ALL: degree 85, zero localization content of any kind in the entry.
LEDGF, degree 60, NO LOCALIZATION RECORD - and its localization is the mechanism it is cited for.
- related to →WDR5 — the partner that is not a partner: it binds the half of KMT2A the fusion throws away
DEGREE 39, NO LOCALIZATION: its only nuclear-compartment mentions are nucleolar stress as a drug phenotype.
SAME SHAPE, DIFFERENT FIELD: a dimension the brief asks for, absent in aggregate, invisible entry by entry, and unenforced by any checker.
ANOTHER GAP VISIBLE ONLY ACROSS ENTRIES rather than within one.