Only 2 of this graph's 26 bounded absences name a deposited dataset — the corpus asks whether a paper exists and almost never whether the data does
also called measured but not reported · absence taxonomy · the deposited-data blind spot
When we say nobody has done an experiment we mean nobody has published a paper, and the measurement may already sit unexamined in a public archive. Only two of our twenty-six such claims check, and the difference matters because one kind of gap needs years and funding while the other needs an afternoon.
The same thing, for a scientist
Enumerated all 26 `absence-*` entries in the corpus on 2026-09-21 and extracted every deposited-data accession pattern (GSE, PXD, PRJNA, SRP, EGAS, phs, MSV) from their full JSON. **Exactly 2 carry one**: `absence-no-menin-inhibitor-occupancy-dataset-in-an-infant-derived-kmt2a-model` (GSE83671) and `absence-no-study-has-tested-for-gained-fusion-occupancy-under-menin-inhibition` (GSE127508). The other 24 bound the absence by PubMed query alone. A PubMed bound establishes that nobody has REPORTED a result; it is silent on whether anyone has MEASURED it, and the two are separated by the entire contents of GEO, PRIDE, SRA and dbGaP. The distinction is not academic: an unmeasured quantity needs a cohort, funding and years, while an unreported one needs somebody to open a supplementary table or re-search a raw file. Tested the sharpest case immediately — `absence-no-kmt2a-fusion-junction-peptide-reported-in-any-mass-spectrometry-dataset` already states in its own data block `not_swept: PRIDE/ProteomeXchange raw files from SEM, RS4;11, MV4-11 re-searched against a fusion-augmented database`, and names no accession. A PRIDE v3 API search returns those datasets, including **PXD009338**, an affinity purification in which an MLL fusion is the BAIT.
Why the two kinds of gap should never share a sentence
| counted on | 2026-09-21 |
|---|---|
| method | enumerated data/graph/nodes/*/absence-*.json, serialised each entry to JSON and regex-matched the accession patterns GSE[0-9]{4,7}, PXD[0-9]{6}, PRJNA[0-9]{5,9}, SRP[0-9]{6,9}, EGAS[0-9]{11}, phs[0-9]{6}, MSV[0-9]{9} |
| absence entries total | 26 |
| entries naming a deposited accession | 2 |
| entries naming none | 24 |
| why the default runs the wrong way | A PubMed bound is cheap, scriptable and produces a defensible number, so it is what gets written. An archive check requires knowing which archive, guessing the right search terms and reading metadata, and it produces a messier answer. The corpus has therefore optimised for the bound it can compute rather than the bound it needs — and every one of the 24 is conservative in the same direction, overstating how empty the world is. |
| shown for fusion | NOT APPLICABLE — this is a corpus-methodology measurement, not a biological claim about any fusion. |
the two that do
{
"absence-no-menin-inhibitor-occupancy-dataset-in-an-infant-derived-kmt2a-model": "GSE83671",
"absence-no-study-has-tested-for-gained-fusion-occupancy-under-menin-inhibition": "GSE127508"
}proposed taxonomy
{
"never_measured": "No instrument has ever been pointed at the question. Closing it needs new material — a cohort, a mouse, a wet experiment. Example: no menin-inhibitor occupancy dataset exists in an infant-derived KMT2A model, because no such model was ever treated and profiled.",
"measured_not_reported": "The assay was run and generates the quantity as a matter of course, but the number was never written down in the text. Closing it needs somebody to open a supplementary table. Example: IDH status in the 36-patient infant KMT2A-r WGS/WES cohort.",
"measured_not_analysed": "The data are deposited and public, but the specific analysis was never run on them. Closing it is compute, not wet work. Example: gained fusion peaks under menin inhibition in GSE127508, where every published pipeline tabulated losses only.",
"measured_not_searched": "The raw signal is in a public archive but could not have been found, because the search space excluded it. Closing it is a re-analysis with a corrected database. Example: a fusion junction peptide is unassignable against a standard human proteome, in which the fusion-spanning sequence has no entry.",
"measured_not_curated": "Published in a paper, but absent from the database this project draws on, so structurally invisible to our joins. Example: PMID 29777171, which IntAct has not entered."
}- This entry supports no clinical action and makes no biological claim. It is an audit of how this project writes its own absence statements.
- It should not be read as saying the graph's absences are wrong. All 24 literature-only bounds are conservative in the same direction: they may understate what is known, and none of them overstates it. No claim elsewhere in the graph becomes less safe because of this finding.
- Add an `absence_class` field with the five-value controlled vocabulary in data.proposed_taxonomy to all 26 absence entries, and require it in the brief. Four entries already carry an ad-hoc absence_class string with free-text values, so the field exists and is unconstrained.
- For each of the 24 literature-only absences, run one archive query in the matching repository — GEO for occupancy, PRIDE for proteomics, dbGaP or EGA for cohort genotypes — and record the result whether positive or negative. Negative archive checks are worth recording precisely because they upgrade a weak absence to a strong one.
- Should the validator require that any entry tagged `bounded-absence` states which class it is? That converts a methodological preference into an enforced one, in the way the affinity-conditions checker did for bare Kd values this cycle.
- Estimate how much this has cost. If a meaningful fraction of the 24 turn out to be measured-but-not-analysed, then the project's standing 'what is next' lists have been recommending wet experiments where a desk job would do.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
Of 26 absence-* entries in this corpus, exactly 2 name a deposited-data accession anywhere in their JSON: GSE83671 and GSE127508. The remaining 24 bound their absence by literature query alone, which establishes non-reporting rather than non-measurement.
Corpus audit of data/graph/nodes/*/absence-*.json, 2026-09-21: 26 entries enumerated, full JSON regex-matched against seven deposited-data accession patterns (GSE, PXD, PRJNA, SRP, EGAS, phs, MSV); 2 matched. · This repository (scripted enumeration; counts and matching entry ids recorded in data.the_two_that_do) 2026What this evidence cannot show
Counts accession STRINGS, not intent. An entry could discuss a public dataset in prose without quoting an accession and would be scored as naming none; conversely quoting an accession does not prove the entry checked it. The pattern list is also not exhaustive — ArrayExpress (E-MTAB), EGA dataset ids (EGAD), MassIVE variants and Zenodo DOIs are not matched, so 2 is a floor. Spot-reading the 24 did not reveal a prose-only case, but 24 entries were skimmed, not read in full.
Evidence that argues against this (1)
- A literature-only bound is not a defect when the question is genuinely about the literature — 'no trial has reported a response rate in NUP98-rearranged patients' is a claim about reporting, and a data archive cannot answer it. So the correct target is not 26 of 26; several of the 24 are properly literature-bounded. The honest finding is that the corpus does not currently DISTINGUISH the two cases, not that all 24 are wrong. This repository 2026
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
The sharpest case resolved on first contact. The junction-peptide absence entry already records in its own data block that PRIDE/ProteomeXchange raw files from SEM, RS4;11 and MV4-11 were 'not swept', and names no accession. A PRIDE v3 API keyword search on 2026-09-21 returns such datasets — including PXD043920 (MLL-AF4 with PAF1 and FACT), PXD041875 (an MLL-AF4 target gene in infant MLL-rearranged B-ALL), PXD042249 (fetal- versus adult-derived MLL-rearranged leukaemia proteome), and PXD009338, 'Protein Interactions of 7 MLL-fusion Proteins', in which an MLL fusion is the affinity-purification BAIT rather than an incidental prey.
PRIDE Archive REST API v3, project keyword search, queried 2026-09-21: keywords 'MLL-AF4' (3 projects), 'KMT2A' (20), 'MV4;11' (13), 'MLL rearranged leukemia' (5), 'RS4;11' (2). · PRIDE Archive / ProteomeXchange (EMBL-EBI) 2026What this evidence cannot show
Keyword search over project titles and metadata only; no raw file was downloaded and no re-search was performed. Existence of a dataset from a KMT2A-rearranged line does not establish that a junction peptide is present or detectable in it — that depends on digestion, coverage and the search database. What this shows is that the sweep the entry itself called for is possible and has named targets, not that it would succeed.
Evidence that argues against this (1)
- Finding candidate datasets is the easy half and it is not the same as closing the absence. A re-search against a fusion-augmented database could return nothing, and a negative there would be genuinely informative only if coverage of the junction region were first shown to be adequate. Recording 'the data exist' as though it were 'the answer exists' would be precisely the over-reading this entry warns against. PRIDE Archive / ProteomeXchange (EMBL-EBI) 2026
Generalises the distinction that entry drew. It closed by asking how many absences recorded as 'nobody has looked' are actually 'nobody has published the cross-tabulation'; the sweep answers that 24 of 26 never asked, naming no deposited accession at all.
Uses it as the worked case: the entry itself states its raw data were not swept and names no archive, and a PRIDE v3 keyword search returns candidate datasets from SEM, RS4;11 and MV4-11 immediately.
One of only two absences in the corpus that names a deposited accession (GSE127508) — cited here as the positive example of the measured-not-analysed class, where closing the gap is compute rather than wet work.
The sweep is what surfaced the AP-MS survey: chasing the junction-peptide entry's unswept-raw-data note into PRIDE returned PXD009338, whose paper turned out to be uncited here and uncurated by IntAct.
A clean measured-not-reported case found by asking that entry's question: both variables are in one trial database on the same 1,022 children, each already published against outcome, and only the cross is missing. Closing it needs permission, not a cohort.