A GEO cell-type field can name the patient's clinical course rather than the aliquot's exposure, and the two readings differ by an entire control arm
Public gene-data records describe each sample in a free-text line. Sometimes that line describes what happened to the patient overall, not what was done to the particular tube of cells - and mistaking one for the other can make a before-and-after experiment look like it has no 'before'.
The same thing, for a scientist
`!Sample_characteristics_ch1` is free text with a curator-chosen key. In GSE252378 the key is `cell type` and the value is a sentence about the donor: `infant B-ALL lineage-switching sample 148752 treated with the Menin inhibitor`. It is applied unchanged to every aliquot from that donor, including the diagnostic aliquot drawn before the drug. Read as an exposure annotation it deletes the study's own control arm. The general form: **a per-sample metadata field can carry a per-PATIENT fact, and nothing in the record marks which it is.** The tell here was an adjacent `!Sample_description` saying `(paired with 152985)` - a pairing statement is a hint that the two aliquots differ in something the characteristics field is not reporting.
Why this is not the same as the off-by-one
This project already records a GEO parsing trap: !Sample_title is emitted before !Sample_geo_accession in the brief format, so a parser keying on the accession line attributes each title to the previous record. That is a mechanical defect and a parser fix closes it. This one is semantic: the parse is correct, the field is correctly read, and the field means something other than what its position in a per-sample block implies. No parser change can detect it.
The three cheap checks
- If two samples are described as paired, ask what differs between them and check whether the characteristics field reports that difference. Here it does not.
- Read the publication's Methods sample-description paragraph before trusting any treatment word in GEO. One paragraph settled this.
- Check whether the string is identical across aliquots that must differ. A treatment annotation that is byte-identical on a diagnostic and a relapse aliquot is describing the patient.
Exposure in this project
This corpus holds 29 entries citing PMID 39472576 and at least one of them - the king entry that surfaced this deposit - describes both aliquots as drug-exposed. Any sweep of GSE252378 that keys on the treatment word in !Sample_characteristics_ch1 is wrong for the four diagnostic aliquot sets (148752 KMT2A-ENL; and by the same logic p279, R3595 and R3652, whose paired relapse samples are p318 and R3652/R3595 respectively).
| field | !Sample_characteristics_ch1 |
|---|---|
| series | GSE252378 |
| failure mode | a per-sample field carrying a per-patient fact |
| cost if missed | the study's within-patient drug-free arm disappears |
| detection | identical treatment string on aliquots that must differ; a 'paired with' note in !Sample_description |
| not the same as | the title/accession off-by-one in the GEO brief format, which is mechanical |
| status | run |
- A metadata-driven cohort builder will silently assign diagnostic aliquots to a treated arm, which makes a drug look ineffective (because the 'treated' baseline already contains the untreated cells) rather than making it look effective.
- The error is invisible to every schema check, because the field is present, well-formed and true.
- How often across GEO does a per-sample characteristics field carry a patient-level treatment statement? Nobody in this project has measured it and the claim here is bounded to one series.
- Should this project's deposit index record, per series, whether treatment annotations are aliquot-level or patient-level, as a field rather than as prose?
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
In GSE252378 the per-sample cell-type field carries a patient-course statement, identical on aliquots drawn before and during menin-inhibitor treatment.
GSE252378_family.soft.gz - the per-sample SOFT records for GSE252378, read raw with line numbers · Janssens DH et al. · NCBI GEO FTP 2024What this evidence cannot show
One series, one curator, one submission. Whether this is common practice across GEO was not measured here and the claim is bounded to this series.
Evidence that argues against this (1)
- GEO's own submission template asks for characteristics as key-value pairs describing the SAMPLE, so a submitter using it for a patient-level fact is departing from the template rather than following it - which argues this is a submitter habit rather than a structural property of GEO, and may be rarer than this entry implies. Janssens DH et al. 2024
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The publication's Methods paragraph resolves the ambiguity in one reading and names the diagnostic and relapse aliquots.
MLL oncoprotein levels influence leukemia lineage identities · Janssens DH, Duran M, Otto DJ, Wu W, Xu Y, Kirkey D, Mullighan CG, Yi JS, Meshinchi S, Sarthy JF, Ahmad K, Henikoff S · Nature Communications 15:9341 2024What this evidence cannot show
Only checkable when the paper is open access and its Methods describes the samples individually. Many depositions have neither.
Evidence that argues against this (1)
- For seven patient samples in this same series the raw data are withheld and only aligned files are shared, so for those the paper's Methods is the ONLY route to the sample's provenance and there is no independent check at all. Janssens DH et al. 2024
the general form of the specific correction