Oncofusion

start / graph / ana-this-corpus-cites-504-deposits-and-a-paper-index-cannot-see-any-of-them

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.

Indexing every deposit this corpus cites - 504 accessions, 102 GEO series, all 102 resolving - because the citation index repaired last cycle indexes papers only, and half the known absence failures were refuted by a dataset rather than by a paper

also called the deposit index · 504 accessions, 102 series, 100% resolving · search the deposits too · the half the citation index could not see

Last cycle this project fixed the tool it uses to check whether something has already been done, which searches every paper its notes cite. That tool cannot see datasets — and one of the two mistakes that prompted the fix was refuted by a dataset, not a paper. This builds the matching index over datasets. It immediately found a public experiment the project had recorded as missing.

The same thing, for a scientist

**Statement.** Cycle 120 repaired the paper citation index and its own verdict named what it still could not do: > *"WHAT WOULD ACTUALLY WORK: index the GEO/SRA accessions this corpus cites and match absences against DEPOSITS, because both known failures were refutable from material the project already held and one of them was a deposit."* **This project has written that kind of sentence six times and executed it once.** This executes it. | | | |---|---:| | distinct accessions cited | **504** | | GEO series | **102** | | resolving to a real deposit | **102 (100%)** | | GSM, PXD, PRJNA, SRP, SRR, E-MTAB | 379, 14, 5, 2, 1, 1 | | series with cached sample-level text folded in | **32** | **`D1` fires and the index paid for itself on its first run.** Matching the 60 bounded-absence entries against deposit text surfaced **GSE252378** as a candidate for `absence-no-menin-inhibitor-occupancy-dataset-in-an-infant-derived-kmt2a-aff1-line`. Reading it found **menin CUT&RUN in two primary infant patient samples under a menin inhibitor** — the experiment that absence says does not exist. It survives only because both patients are `KMT2A-ENL` rather than the index fusion. **That deposit was cited by one entry in this corpus and no amount of reading entry prose would have surfaced it.** **`U2` is the binding limit and it is not small.** A GEO **series title describes the submission, not every arm in it.** GSE159608's series title is *"Automated CUT&Tag profiling of chromatin heterogeneity in mixed-lineage leukemia"* — it does not mention SEM, VTP-50469 or a menin inhibitor, all of which live in its **sample** records. **So series-level matching would still have missed the known failure**, and sample-level text is folded in wherever this project already has it cached — which is 32 of 102 series. **The other 70 are indexed at series level only and the index is correspondingly blind in them.** **THE RULE, now complete:** before a brief claims an experiment has never been done, search the **paper** index *and* the **deposit** index. Cycle 118 built the first and got it wrong; cycle 120 fixed it; this is the second half.

Go deeper
Why 100% resolution means something different here than it did last cycle

Cycle 118 reported 100% of cited PMIDs resolve and that number turned out to be a precision measure over a quarter of the corpus. The same trap is available here and is handled explicitly.

D3 was pre-registered: an HTTP 200 carrying a body without !Series_title is a FETCH FAILURE, not an empty deposit, and must never be scored as no match. That distinction is the deposit-index version of the empty-querytranslation rule this project learned on E-utilities. All 102 returned a real record, so the clause did not have to fire — but it exists, and the 100% is a statement about fetching, not about coverage of the literature.

Recall is bounded by U3 instead: this corpus cites 102 GEO series and GEO holds hundreds of thousands. The index bounds what this project holds, which is exactly the question a dispatch brief needs and nothing more.

What the accession census says about where this corpus rests
accessionentries citing it
GSE260742147
GSE12750875
GSE12750773
PXD00933856
GSE21048049
GSE12750641

One deposit is cited by 147 entries. For comparison the most-cited paper in this corpus appears in 29. This graph rests far more heavily on a handful of deposits than on any paper, which is worth knowing before the next re-analysis beat is chosen — and it means a defect in one deposit propagates much further than a defect in one citation.

🔴 That is not a hypothetical. Cycle 120 measured that 16.4% of GEO sample labels contradict their own metadata, and GSE159608 alone supplied 55 of 96 contradictions. A deposit cited by 147 entries whose per-sample labels are unreliable is a much larger exposure than any single wrong PMID, and no entry in this graph currently records the label-integrity status of GSE260742.

Facts and figures
questionwhat deposits does this corpus cite, and can they be searched before claiming a gap?
distinct accessions504
geo series resolved102 of 102 (100%)
series with cached sample text32
absence entries scored60
candidates at 0 50
candidates 0 3 to 0 54
for comparison most cited paper29
what it found on its first runGSE252378 - menin CUT&RUN in two primary infant patient samples under a menin inhibitor, cited by one entry here, covering an absence this project had asserted; the absence survives only because both patients are KMT2A-ENL
THE RULEbefore a brief claims an experiment has never been done, search the PAPER index and the DEPOSIT index; cycle 120 fixed the first half, this is the second
the binding limitU2 - a series title describes the submission, not every arm; the known failure lived in SAMPLE records, and sample text is cached for only 32 of 102 series
unmeasured exposureGSE260742 is cited by 147 entries and no entry records its label-integrity status, although 16.4% of GEO sample labels were measured to contradict their own metadata last cycle
absence classmeasured-not-curated
bite backThis index makes 'has this been deposited?' cheap to answer and therefore cheap to assert. The previous index had the same property and was used to justify a novelty claim that was false fourteen times over. If briefs start citing this index as evidence of a gap rather than using it to avoid one, it will have automated the habit rather than the check.
by kind
SRR1.000SRR: 1E-MTAB1.000E-MTAB: 1SRP2.000SRP: 2PRJNA5.000PRJNA: 5PXD14.0PXD: 14GSE102.0GSE: 102GSM379.0GSM: 379
most cited deposits
GSE12750641.0GSE127506: 41GSE21048049.0GSE210480: 49PXD00933856.0PXD009338: 56GSE12750773.0GSE127507: 73GSE12750875.0GSE127508: 75GSE260742147.0GSE260742: 147
by kind — the numbers behind the chart
{
  "GSE": 102,
  "GSM": 379,
  "PXD": 14,
  "PRJNA": 5,
  "SRP": 2,
  "SRR": 1,
  "E-MTAB": 1
}
most cited deposits — the numbers behind the chart
{
  "GSE260742": 147,
  "GSE127508": 75,
  "GSE127507": 73,
  "PXD009338": 56,
  "GSE210480": 49,
  "GSE127506": 41
}
Still open
  1. Fetch sample-level records for the 70 GEO series that have none cached. U2 is the binding limit and this closes most of it; the known failure lived in sample records and the index is currently blind inside 70 of 102 series.
  2. Audit GSE260742's label integrity. It is cited by 147 entries, more than five times the most-cited paper, and 16.4% of GEO sample labels were measured to contradict their own metadata last cycle. No entry records whether this one does.
  3. Measure this index's RECALL, rather than reporting its fetch rate. The previous index reported 100% resolution while missing 73.7% of its corpus, and nothing here has tested whether the accession regex misses deposits cited by URL or by name.
  4. Extend to PRIDE, dbGaP and EGA. 14 PXD accessions are counted but never resolved, and controlled-access repositories match no pattern here at all.
Every source behind this entry (1)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    Extracting every string matching GSE, GSM, SRP, SRR, PRJNA, PXD or E-MTAB accession patterns from all 2,633 committed node files yields 504 distinct accessions: 102 GEO series, 379 GEO samples, 14 PRIDE, 5 BioProject, 2 SRA studies, 1 SRA run and 1 ArrayExpress. Fetching each of the 102 series from the GEO accession endpoint returns a record carrying a !Series_title for 102 of 102, 100.0%, with no fetch failures. Cached sample-level text exists in this repository for 32 of the 102 and is folded into the searchable text for those. Scoring the 60 bounded-absence entries against deposit text by subject-vocabulary containment, excluding deposits each absence itself cites, gives 0 absences with a candidate at 0.5 or above and 4 between 0.3 and 0.5. The most cited accession, GSE260742, appears in 147 entries.

    What this evidence cannot show

    A SERIES TITLE DESCRIBES THE SUBMISSION, NOT EVERY ARM IN IT, and this is the binding limit rather than a footnote: GSE159608's series title names neither SEM nor VTP-50469 nor a menin inhibitor, all of which are in its SAMPLE records, so series-level matching WOULD STILL HAVE MISSED the known failure this index was built to catch. Sample-level text is available for only 32 of 102 series and the index is blind inside the other 70. A DEPOSIT MATCH IS A CANDIDATE, NEVER A REFUTATION - GSE252378 covers the infant menin-inhibitor question and does not refute the absence, because the fusion differs, and only reading settled that. THE 100% IS A FETCH RATE, NOT A COVERAGE RATE: this corpus cites 102 GEO series against hundreds of thousands in GEO, so absence of a match bounds what the PROJECT holds and says nothing about the literature. The accession regex will miss deposits referenced by URL or by name without an accession, and dbGaP, EGA and controlled-access repositories are not covered by any pattern here.

    Evidence that argues against this (3)
    • The 0-at-0.5 result looks like a clean bill of health for the absence claims and is not one. The companion analysis returned an explicitly UNVALIDATED verdict for exactly this reason, and the one deposit-refutable absence this index did surface scored 0.333 - below any threshold that would have flagged it automatically. It was found by reading the worklist, not by the score. An index whose useful output requires a human to read past the ranking has not automated the check it claims to automate.
    • Building a second index is the same response that produced the first one, and the first one was wrong for two cycles while looking authoritative. This index has been run once, by its author, on the afternoon it was written, and its 100% fetch rate is precisely the kind of confident-sounding number that concealed the previous defect. Nothing here has measured its RECALL against deposits the corpus cites in a form the regex does not match.
    • The deeper problem may not be indexable at all. Four consecutive briefs asserted a gap that did not exist, and the response each time has been a better search tool. A brief that states what is known and asks for an extension cannot fail this way regardless of index quality, and that change was proposed in cycle 118, repeated in cycle 120, and still has not been made a rule.
Connected to