Oncofusion

start / graph / ana-the-deposits-agree-with-themselves-on-1419-pairs-and-five-cannot-be-checked-at-all

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

Re-auditing every cached GEO deposit on three axes after the parser fix: 1,597 comparable title-versus-characteristics pairs across 40 series and 2,620 samples, zero disagreements - and nine series including the one this cycle's beats rest on are abstentions rather than passes, because a 0-of-0 is not a clean bill of health

also called the deposits agree with themselves · 1,419 pairs, zero disagreements · a 0-of-0 is an abstention not a pass · GSE260742 records no antibody anywhere · the cell-line axis doubled the audited surface

A previous entry claimed public data archives disagree with their own descriptions about one time in six. That was a bug in our own code, now fixed. Re-checking properly, and on a third kind of label as well, they agree every single time across fourteen hundred checks. The more useful finding is the opposite of a defect: five archives cannot be checked at all, including the one the project leans on most, and 'nothing found' there means 'nothing looked at'.

The same thing, for a scientist

**Statement.** After the off-by-one retracted in cycle 121, the deposit self-consistency audit was re-run and extended to a third axis. | axis | comparable pairs | disagree | |---|---:|---:| | antibody | 423 | **0** | | treatment | 357 | **0** | | **cell line** (new) | **817** | **0** | | **total** | **1,597** | **0** | **`A1` fires: the deposits agree with themselves.** Across 40 series and 2,620 samples there is not one title-versus-characteristics contradiction. 🟠 **Those totals moved three times in one session, upward, because I twice widened the instrument after seeing what it could not see** — first a cell-line axis, then the standard histone marks. 666 → 1,419 → 1,597 comparable pairs. Each addition is defensible in isolation (`H3K9me2` is plainly an antibody) and **none was pre-registered**, so the growth of the denominator is a record of my attention rather than of the archive. The cycle-120 claim of 16.4% was entirely a parser defect and the seven residuals after the fix were entirely a tokeniser defect. **The cell-line axis more than doubled the audited surface, and it was added for a specific reason.** `GSE260742` is **cited by 147 entries — five times the most-cited paper in this corpus** — and it scored **0 of 0** on antibody. **A 0-of-0 is an ABSTENTION, not a pass**, and printing it beside genuine passes is how the single largest deposit in this graph came to look audited when it was unexamined. Its characteristics carry `tissue`, `cell line`, `cell type` and `treatment: 7 days` — **never an antibody.** **On the new axis GSE260742 is clean: 0 of 37.** But the consequence for the corpus is the opposite of reassuring: **its characteristics record no antibody for any of its 158 samples, so the antibody assignment for every one of them rests on the sample title alone, with no second field to check it against.** For a deposit 147 entries depend on, that is the exposure worth recording. **Nine of 40 series are unexaminable by this instrument**, listed so nobody mistakes their silence for a pass: `GSE252378` (**171 samples — the deposit both of this cycle's beats rest on**), `GSE129524` (48), `GSE120221` (25), `GSE192561` (24), `GSE121740` (18), `GSE129526` (12), `GSE214158` (12), `GSE189105` (8), `GSE294098` (2). 🔵 **Eight series were fetched this cycle specifically to close this gap**, chosen as the most-cited deposits with no cached sample text. It worked for one that mattered: **`GSE210480` — 60 samples, cited by 49 entries — moved from unexaminable to audited** once the standard histone marks were in the vocabulary, because it records `antibody: H3K9me2` and titles its samples `..._H3K9ME2_rep1`. **It had been invisible purely because my token list stopped at the marks this project happens to study.** 🔵 **The generalisable rule this earns:** an audit must report **what it could not look at** beside what it found. This is the same lesson as cycle 120's citation index — a precision measure with no recall measure — arriving at a different instrument three cycles later.

Go deeper
Why GSE252378 shows as unexaminable here when a purpose-built parse found it clean

Cycle 121 reported 0 of 44 antibody disagreements in GSE252378 using an ad-hoc parse written for that deposit. This generic audit reports 0 of 0 comparable pairs for the same series. Both are correct and the difference is the vocabulary.

That deposit names its antibodies as Rabbit Menin Bethyl Labs A300-115A, Mouse KMT2A-N1 Millipore 05-764 and so on. The generic list here carries MENIN and MLL1 but not KMT2A-N1/N2/C1/C2 as distinct tokens, so no pair is scored as comparable.

🟠 This is worth stating rather than quietly reconciling, because it shows the audit's coverage is a property of the coordinator's vocabulary list and not of the deposits. Every 0 of 0 in the per-series table is a statement about my token lists. Extending them would move series from the abstention column into the audited column, and nobody has measured how many.

What a clean result does and does not license

It licenses taking sample titles at face value in these 32 series — which matters immediately, because cycle 120 used a false contradiction rate to discount GSE159608's SEM DMSO / EPZ-5676 / VTP-50469 arm labels, and an agent then confirmed those labels independently from the signal (28 of 41 libraries nearest-neighbour to the same antibody).

It does not license believing the deposits are correct. A1 says title and characteristics agree; both are written by the same submitter and can be wrong together. The only check that would catch a jointly-wrong label is the signal, and that has been done for exactly one series.

🔴 And a zero is the easiest number to over-read. The two normalisations that took the residual count from 7 to 0 in cycle 121 were written after seeing which cases failed. Both are unambiguous on the raw records — a 0 µM dose is a vehicle, and MLL1 (c terminus) is the same antibody as MLL1c — but the sequence was 16.4% → 1.1% → 0% with the instrument edited between the last two steps, and that is an uncontrolled degree of freedom.

Facts and figures
questionafter the parser fix, do the cached GEO deposits agree with their own metadata?
verdictA1 - they agree; 0 disagreements in 1,419 comparable pairs
series40
samples2620
total comparable1597
total disagree0
supersedesthe retracted 16.4% claim, which was an off-by-one in this script's own record boundary, plus seven tokeniser artefacts
why the cell line axis was addedGSE260742 scored 0 of 0 on antibody, and a 0-of-0 is an abstention rather than a pass; its titles and characteristics both name the cell line
absence classmeasured-not-curated
bite backThis entry reports zero contradictions in 1,419 checks, which reads as a clean archive. The same instrument reported 16.4% four days ago with equal confidence. What changed is a record boundary and two synonyms, and nothing external has confirmed the new number either - a reader should hold it more loosely than its precision suggests.
the denominator grew three times666 -> 1,419 -> 1,597 comparable pairs, none pre-registered
antibody
{
  "comparable": 423,
  "disagree": 0
}
treatment
{
  "comparable": 357,
  "disagree": 0
}
cell line
{
  "comparable": 817,
  "disagree": 0
}
abstentions not passes
{
  "GSE252378": 171,
  "GSE129524": 48,
  "GSE120221": 25,
  "GSE192561": 24,
  "GSE121740": 18,
  "GSE129526": 12,
  "GSE214158": 12,
  "GSE189105": 8,
  "GSE294098": 2,
  "note": "no comparable pair on any axis; these are UNEXAMINED, and GSE252378 is the deposit both of this cycle's beats rest on"
}
the most cited deposit
{
  "accession": "GSE260742",
  "cited_by_n_entries": 147,
  "samples": 158,
  "antibody": "0 of 0 - its characteristics never name an antibody",
  "treatment": "0 of 2",
  "cell_line": "0 of 37 - clean",
  "the_exposure": "the antibody assignment for all 158 samples rests on the sample title alone, with no second field to check it against, in a deposit 147 entries depend on"
}
sample records fetched this cycle
{
  "count": 8,
  "chosen_as": "the most-cited GEO series with no cached sample text",
  "accessions": [
    "GSE210480",
    "GSE139196",
    "GSE225028",
    "GSE299100",
    "GSE121740",
    "GSE129526",
    "GSE120221",
    "GSE129524"
  ],
  "outcome": "GSE210480 - 60 samples, cited by 49 entries - moved from unexaminable to audited once standard histone marks were in the vocabulary; four of the eight are still abstentions",
  "why": "the deposit index was blind inside 84 of the 102 GEO series this corpus cites"
}
Still open
  1. Extend the antibody vocabulary to the forms deposits actually use - 'Rabbit Menin Bethyl Labs A300-115A', 'Mouse KMT2A-N1 Millipore 05-764' - and re-measure how many of the NINE abstaining series become auditable. The surface already grew from 666 to 1,597 pairs through two unplanned extensions, so this dependence is large and unmeasured.
  2. Set a threshold below which a series counts as effectively unexamined. GSE260742 has 2 comparable treatment pairs across 158 samples and is currently counted as audited; that is not meaningfully different from an abstention.
  3. Check a second series at SIGNAL level. Agreement between two fields written by the same submitter cannot catch a jointly-wrong label, and exactly one series has had its labels confirmed from the data.
  4. Record the label-integrity status of every deposit an entry depends on, not just the cached 32. The corpus cites 102 GEO series and this audit covers 32 of them.
  5. Fetch sample records for the remaining 76 cited GEO series with none cached. Eight were fetched this cycle and one of them mattered; 84 of the 102 cited series were blind and 76 still are.
Every source behind this entry (1)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    Parsing every cached GEO sample record across 40 series and 2,620 samples with record boundaries taken from !Sample_title, and comparing each sample's title against its own characteristics and treatment_protocol fields on normalised antibody, treatment and cell-line vocabularies: 423 comparable antibody pairs with 0 disagreements, 357 comparable treatment pairs with 0 disagreements, and 817 comparable cell-line pairs with 0 disagreements, 1,597 comparable pairs in total and no contradiction on any axis. Nine series have no comparable pair on any axis and are unexaminable: GSE252378 with 171 samples, GSE129524 with 48, GSE120221 with 25, GSE192561 with 24, GSE121740 with 18, GSE129526 with 12, GSE214158 with 12, GSE189105 with 8 and GSE294098 with 2. GSE260742, which 147 node files cite, has 0 comparable antibody pairs because its characteristics never name an antibody, 0 of 2 on treatment and 0 of 37 on cell line. GSE210480, cited by 49 entries, moved from unexaminable to audited when standard histone-mark antibodies were added to the vocabulary.

    What this evidence cannot show

    COMPARABILITY IS A PROPERTY OF MY TOKEN LISTS, NOT OF THE DEPOSITS, AND THIS WAS DEMONSTRATED TWICE IN ONE SESSION: the audited surface went 666 to 1,419 to 1,597 comparable pairs as I added a cell-line axis and then the standard histone marks, each time after seeing what the instrument could not see, and neither extension was pre-registered. GSE252378 scored 44 comparable pairs under a purpose-built parse and 0 under this generic one because it names antibodies as 'Rabbit Menin Bethyl Labs A300-115A'. How many of the nine abstaining series would become auditable under a fuller vocabulary has not been measured. THE DENOMINATOR IS COMPARABLE PAIRS, NOT SAMPLES: 1,597 pairs against 2,620 samples. AGREEMENT IS NOT CORRECTNESS - title and characteristics are written by the same submitter and can be wrong together; only signal-level checking would catch that and it has been done for one series. THE ZERO WAS ALSO REACHED BY EDITING THE INSTRUMENT: two tokeniser normalisations took the residual from 7 to 0. NOTHING HERE OPENS A LIBRARY.

    Evidence that argues against this (3)
    • A clean sweep produced by the author of the retracted claim, on the instrument that produced it, days later, is the least independent confirmation available. The only external check on any of this is one agent's signal-level clustering of one series, which agreed with the titles; the other 31 series rest on the corrected script alone.
    • Adding an axis that happens to make the most-cited deposit auditable, after finding it scored 0 of 0, is a choice made with the result in view. The cell-line axis is defensible on its own terms, but it was not pre-registered, and a reader should note that the audited surface grew from 666 to 1,419 pairs through a decision taken because the previous number was unsatisfying.
    • The abstention list is itself incomplete in a way that flatters the instrument. It counts only series with zero comparable pairs on ALL THREE axes; a series with two comparable pairs out of 158 samples, as GSE260742 has on treatment, is counted as audited rather than as effectively unexamined, and no threshold separates the two.
Connected to