When a paper will not say what it was done in, ask its data: resolving an unstated experimental system from GEO
also called GEO provenance lookup · bounded-unreadable escape hatch
If a paper is behind a paywall and its summary does not say which cells the experiment used, you can usually find out for free by looking at the sequencing data the authors had to deposit, which lists the species and cell type for every sample.
The same thing, for a scientist
A recurring failure in this graph is an entry that records *'the source does not state its experimental system'* as a caveat and stops. That is an honest sentence and a terminal one: it closes a question that is usually answerable. Sequencing data deposition is a condition of publication at essentially every journal in this field, and GEO records **organism** and **cell type** as per-sample fields, together with growth and treatment protocols. **The route, in three calls.** (i) `esearch db=gds` on the paper's distinctive method name, its first author, or its title words. (ii) `esummary db=gds` on the returned UIDs to identify the series and its taxa. (iii) the accession-display text endpoint, `https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=<GSE>&targ=all&form=text&view=brief`, which returns `!Sample_organism_ch1`, `!Sample_characteristics_ch1 = cell type: ...`, `!Sample_title`, and the shared growth and treatment protocol strings. This beat used the route to establish that the CRISPR-ChIP H3K79 partition is a murine MLL-AF9 result after the publisher, Europe PMC, Unpaywall, Semantic Scholar and the bioRxiv preprint had all failed [E1][E2][E3].
What the route can and cannot establish
Can: organism, cell type as the submitters describe it, antibody used, treatment and growth protocols, which assays were run in which system, and how many samples of each - often at a granularity the abstract never reaches.
Cannot: anything not sequenced. A western blot, a ChIP-qPCR, a viability assay or a mouse survival curve in a second system leaves no GEO trace, so a conclusion from GEO is bounded as 'of the deposited sequencing', never as 'of the study'. Nor does it establish which figure rests on which sample.
And it inherits the submitter's accuracy. GEO fields are submitter-entered and unverified; this beat found four samples in one series labelled with the cell types 'K563', 'K564', 'K565' and 'K566', evidently for K562 2. A correction sourced from GEO should say so, so that a downstream reader knows which authority to blame.
Where else the same move works
- ArrayExpress / BioStudies for European depositions.
- SRA BioSample records, which carry organism, cell line, tissue and sometimes karyotype as structured attributes, reachable from any GEO sample.
- Addgene entries for plasmids named in a paper, which often state the construct's exact boundaries when the paper does not.
- ClinicalTrials.gov raw API for
whyStopped,armsInterventionsModuleand collaborator lists, which this project already documents - the same principle applied to a registry rather than a repository.
The common shape: the administrative record of a piece of work is frequently more specific, and always more open, than its narrative record.
Correction: step (i) of this method never worked
This entry prescribed esearch db=gds&term=<PMID>[PMID] as the way in. It returns zero records for every paper tested — including PMID 37679565, the paper this method was written from and whose four GEO series this graph documents in detail.
Verified by the coordinator directly on 2026-09-19:
esearch db=gds term=37679565[PMID] -> count 0, warninglist: ['No items found.']
elink dbfrom=pubmed db=gds id=37679565 -> 4 uids: 200192562 200192561 200192555 200192544
Those four are GSE192562, GSE192561, GSE192555 and GSE192544 — exactly the series the CRISPR-ChIP resolution rests on. The broken step returned a well-formed esearchresult with count: 0 and "No items found.", twenty-six times in a row, which is the most convincing possible presentation of a false absence.
The corrected route has two co-equal entry points, because neither alone is sufficient:
elink dbfrom=pubmed&db=gds&id=<PMID>— returns GDS uids; strip the leading200to get the GSE number.- The paper's data-availability statement. Six of sixteen resolutions in the Q18b sweep rested on accessions found only there; elink alone would have given 37.5% instead of 66.7%.
Then https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=<GSE>&targ=gsm&form=text&view=brief for the per-sample SOFT records.
Two further routes that are unusable as evidence of absence: elink pubmed→sra, →biosample and →bioproject returned 0 for 26 of 26 papers, including several with confirmed GEO and ArrayExpress depositions.
And attribute by !Series_pubmed_id, never by topic. A GEO keyword search on DOT1L AND "O-GlcNAc" returns exactly one series with an almost-perfect topic match, belonging to a different paper. Attributing it would have produced a clean, specific, wrong answer.
| bound | of the deposited sequencing, never of the study |
|---|---|
| broken route do not use | esearch db=gds&term=<PMID>[PMID] - returns count 0 for every paper tested, including PMID 37679565 whose four series elink returns; verified 2026-09-19 |
route
[ "elink dbfrom=pubmed&db=gds&id=<PMID> (strip the leading 200 from each uid to get the GSE)", "the paper's data-availability statement (co-equal: 6 of 16 resolutions came only from here)", "https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=<GSE>&targ=gsm&form=text&view=brief", "attribute by !Series_pubmed_id, never by topic match" ]
fields returned
[ "!Sample_organism_ch1", "!Sample_characteristics_ch1 (cell type, time point, antibody)", "!Sample_title", "!Sample_growth_protocol_ch1", "!Sample_treatment_protocol_ch1" ]
alternatives
[ "ArrayExpress/BioStudies", "SRA BioSample attributes", "Addgene construct records", "ClinicalTrials.gov raw API" ]
- Treating GEO as authoritative imports the submitters' errors into the graph with higher confidence than the honest unknown it replaces, so every GEO-sourced provenance claim should name GEO as the authority rather than the paper.
- The route makes it cheap to state a system and does not make it cheap to state a result; an agent that resolves provenance from GEO and then writes about the paper's findings has used the escape hatch for the wrong question.
- How many entries in this graph carry a caveat of the form 'the source does not state' that a deposition lookup would resolve? That grep is the work queue this concept implies.
- Is there an equivalent route for papers with no sequencing - proteomics via PRIDE, imaging via IDR, flow via FlowRepository? The principle should hold but this beat did not test it.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
The published CRISPR-ChIP paper's full text could not be obtained through Europe PMC (not open access, no PMCID), Unpaywall (oa_status closed, no OA location), Semantic Scholar (openAccessPdf empty) or the bioRxiv preprint (PDF, .full HTML and JATS XML all returning HTTP 429 with Cloudflare error 1015 across four attempts and two clients).
CRISPR-ChIP reveals selective regulation of H3K79me2 by Menin in MLL leukemia. · Gilan O, Talarmain L, Bell CC, Neville D, Knezevic K, Ferguson DT, Boudes M, Chan YC, Davidovich C, Lam EYN, Dawson MA · Nat Struct Mol Biol 2023What this evidence cannot show
All routes attempted on 2026-09-19 from a single network location. HTTP 429 is a rate limit rather than a permission denial, so the bioRxiv preprint is likely readable from another network or at another time; this is a bounded unreadable for this session, not a permanent one. Institutional subscription access was not available to this agent and was not attempted.
Evidence that argues against this (1)
- A reader with institutional access or a different IP would have read the paper directly and never needed this route, so the method's value is contingent on access rather than intrinsic - which is an argument for recording WHICH routes failed, so a better-resourced reader can skip straight to the paper. Gilan O et al. 2023
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
The GEO accession-display text endpoint returns per-sample organism, cell type, title, antibody and shared growth and treatment protocols for GSE192544 and GSE192555, which together identify every H3K79me2 experiment in the study as Mus musculus, cell type MLL-AF9.
CRISPR-ChIP to identify mechanisms of chromatin regulation [ChIP-seq] (GSE192555) · Gilan O, Dawson MA, et al. (GEO submitters) · NCBI Gene Expression Omnibus 2023What this evidence cannot show
Read on 2026-09-19. Submitter-entered and unverified: four samples in this series carry the cell types 'K563' through 'K566', evidently typographic variants of K562. The endpoint returns the whole series as plain text, which for a large series is tens of megabytes and must be parsed rather than read.
Evidence that argues against this (1)
- Per-sample annotation quality varies enormously between submissions; many series record cell type only in a free-text title, and some record nothing beyond organism, in which case the route returns less than the abstract. Gilan O et al. 2023
- E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
Europe PMC's record for PMID 37679565 reports isOpenAccess N, inEPMC N, no PMCID, and a single full-text URL marked 'Subscription required'; its annotations API returns only 26 abstract-level text-mined terms for the article, none identifying a cell line or organism.
Europe PMC REST search and annotations API records for EXT_ID:37679565 · Europe PMC Consortium · Europe PMC 2026What this evidence cannot show
Retrieved on 2026-09-19. The annotations API mines full text where available and abstracts otherwise; for a subscription article it returns abstract-level terms only, so the absence of a cell-line annotation reflects the access level and not the paper's content.
Evidence that argues against this (1)
- Europe PMC's text-mined annotations are sometimes richer than this for subscription articles where the publisher supplies text for mining, so a negative here should be checked rather than assumed for other papers. Gilan O et al. 2023
- E9established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
The route this entry originally documented does not work. `esearch db=gds` with `<PMID>[PMID]` returns count 0 and 'No items found.' for PMID 37679565, whose four GEO series (GSE192544, GSE192555, GSE192561, GSE192562) are returned immediately by `elink dbfrom=pubmed&db=gds` for the same PMID. The same esearch route returned zero for all 26 papers in the Q18b worklist.
NCBI E-utilities: elink dbfrom=pubmed&db=gds, and esearch db=gds PMID-field behaviour · NCBI E-utilities 2026What this evidence cannot show
Verified by the coordinator directly on 2026-09-19 against the live E-utilities endpoints, after a research agent reported it. The elink uids carry a 200 prefix that must be stripped to give the GSE accession. Why the PMID field behaves this way in db=gds was not investigated; the finding is behavioural, not an explanation.
Evidence that argues against this (2)
- The failure is silent rather than absent, which is the part that matters and also the part that limits the claim. A well-formed `esearchresult` with count 0 is a valid response to a query that finds nothing, so nothing in the output is wrong — the defect is in the query, and only comparing it against a known-positive case reveals that. Any beat that used this route since cycle 49 recorded negatives that look identical to real ones. NCBI E-utilities 2026
- Recording this as a broken route slightly overstates it: the same term syntax works in other E-utilities databases, and an indexing change at NCBI could restore it without notice. The durable lesson is the spot-check, not the endpoint — this was caught only because the method was run against a paper whose answer was already on record. NCBI E-utilities 2026
the worked example
the method that makes provenance auditable
an escape hatch for one class of bounded unreadable
- ← supportsWhat this beat read and did not use: bounded absences, bounded rejections and one bounded unreadable
the route used when the paper could not be read
applies the GEO provenance route to a second unstated system
applies the GEO provenance route
applies the route, and documents a second class of submitter annotation error
a case where the deposition answers the asked question while being the wrong experiment
a case where every listed accession is a re-analysis cohort or unreachable, so the route returns nothing
establishes that a keyword-found series must be attributed by Series_pubmed_id, never by topic
answers this concept's own open question: how many entries carry a caveat of this form
gives the measured prior that makes the GEO route worth running by default
step (i) of this concept's documented route returns zero for all 26 papers tested, including the one it was written from