A GEO deposit can name two different cell lines for the same four samples in two different metadata fields, and reading only the filename - which is what every automated pipeline does - shows you neither: the case where my dispatch brief was right, my correction of it was wrong, and the deposit supports neither of us
also called HCB1 or HCB2 through HCB5 · source_name versus treatment_protocol · the filename shows neither · both of us overstated
A public dataset described four of its samples as coming from one cell line in one field and from four different cell lines in another, while the filenames said something else again. This is worth recording because everyone - people and software alike - reads the filename, and the filename was the one source that could not settle it.
The same thing, for a scientist
**The case.** GSE173599's four MLL1/MLL1c samples carry: | field | what it says | |---|---| | **filename / `!Sample_title`** | `HCB1_ChIP_MLL1c_DMSO...` and `hsCB-FKBP12-MA9_ChIP_MLL1_DMSO_3h...` — **two different naming schemes for the same experiment** | | **`!Sample_source_name_ch1`** | **HCB2, HCB3, HCB4, HCB5** — one per sample, in sequence | | **`!Sample_treatment_protocol_ch1`** | *"human CD34+ MLL-AF9-HA-FKBP12 transformed cells **(HCB1)**"* | | **`!Sample_characteristics_ch1`** | no `cell line:` field at all — unlike the HA samples, which carry `cell line: HCB1` explicitly | **Every other sample in the series has a `source_name` that is a verbatim copy of its own title.** These four are the only ones where `source_name` carries a distinct value, and that value increments by one across the four. **Two readings, and the deposit does not choose between them.** Either HCB2-HCB5 are **four independent cord-blood transformants** — in which case the deposit contains cross-line replication nobody has used, and it is the more interesting possibility — or the submitter filled a spreadsheet column incrementally and these are **all HCB1**, which the treatment protocol asserts. **An incrementing sequence across exactly the four samples that lack a `cell line:` field is suggestive of the second, and suggestive is not resolved.** **THE EPISTEMIC POINT, which is why this is a concept entry.** My dispatch brief said *"and `MLL1` in a second line"*. My own pre-registration, written after checking the filenames, **recorded that the brief was wrong** and that MLL1 was in HCB1 like everything else. **Then the sample-level metadata showed the brief's claim had support the filenames concealed — and also showed that no field in the deposit settles it.** The brief overstated, my correction overstated in the opposite direction, and **the correction felt more reliable because it was a check.** **A check that reads one field is a claim about that field.** The filename is the field every pipeline reads, and here it is the field that carries neither answer.
The operational rule
Before attributing a public sample to a cell line, read source_name, characteristics and treatment_protocol, not the title. They are separate free-text fields filled at different times by different people, and they disagree often enough to matter.
And when they disagree, record the disagreement rather than picking one. This entry's sibling (ana-q139-the-mll1-epitope-specificity-control-inverts-and-is-uninterpretable) reports an inverted result that would have been publishable as a finding if the cell line had been assumed. Naming the ambiguity is what made the refusal to interpret defensible instead of arbitrary.
The generalisation beyond GEO: a correction is a claim and inherits the same duty of evidence as the thing it corrects. Correcting a brief feels like diligence, so it receives less scrutiny than an original assertion — which is exactly backwards when the correction is based on a narrower check than the claim it overturns.
- This entry argues for reading more metadata fields, but more fields also means more opportunities to find a disagreement and declare a result uninterpretable. The rule should be to record disagreements, not to treat any disagreement as disqualifying - a deposit with imperfect metadata is still usable for most questions.
- Are HCB2 through HCB5 real independent transformants? If they are, this deposit contains biological replication across cord-blood donors that no analysis has used, and that would be worth considerably more than the epitope control it was found inside.
- How many other samples this project has analysed carry a source_name that disagrees with their filename? A one-pass check across every GEO sample cited in this corpus is cheap and has never been run.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
GSE173599's four MLL1/MLL1c samples carry source_name values HCB2, HCB3, HCB4 and HCB5 while their own treatment_protocol field states the cells are HCB1, and unlike every other sample in the series they carry no cell-line characteristics field.
GSE173599 sample records GSM5696327, GSM5696328, GSM5696329, GSM5696330, retrieved as ?targ=self&form=text 2026-09-23 · Olsen SN, Armstrong SA, et al. · NCBI Gene Expression Omnibus 2022What this evidence cannot show
Thirteen of the series' samples were pulled individually and compared; the remaining samples were not, so the claim that these four are the ONLY ones with a distinct source_name rests on a thirteen-sample comparison plus the filename pattern across all 54 peak-table samples, not on an exhaustive per-sample pull. The published paper's STAR Methods were read for cell-line description but do not enumerate HCB2-HCB5.
Evidence that argues against this (1)
- If the paper's STAR Methods or supplementary tables enumerate multiple independent HCB transformants, the HCB2-HCB5 reading is correct and this is not a metadata defect but under-used biological replication. That section was not read exhaustively and should be before the ambiguity is quoted as a data-quality problem. Olsen SN et al. 2022
The first of three reasons this arm is not interpreted: the deposit names HCB2-HCB5 in source_name and HCB1 in treatment_protocol for these four samples, and the filenames carry neither answer.
- related to →What is a fusion gene?
A data-provenance caution attached to the fusion-biology entry point: the cell line an experiment used is metadata, it can be recorded inconsistently in a public deposit, and the filename every pipeline reads may carry no answer at all.