label-samples

What this shows

Public sequencing metadata is messy free text. This pipeline fetches six real runs from the ENA, asks Jev to label each one against a fixed vocabulary, and splits the cohort into labels you can trust and labels a human should look at.

The point is the split. Jev returns a probability for each label, so the accept/review decision is made by the pipeline — a number you can compare in nextflow.config — not by asking a model to rate its own confidence.

nextflow run examples/label-samples

What gets asked

One request per sample, carrying that sample's raw ENA record and four questions about it:

question	kind	what comes back
Which organism is this run derived from?	pick one	Homo sapiens, and how sure
Which assay does this run describe?	pick one	RNA-seq, and how sure
Is this an immortalised cell line?	yes/no	a probability, e.g. 0.95
How well evidenced is the tissue of origin?	a level	0–2 on a described scale

In the script those are the choice, noul and score builders — the three kinds of question Jev answers. All four are sent together and answered against the same metadata, so a sample costs one call, not four.

Nothing else is sent: only the ENA record, which is already public.

What you should see

ACCEPT SRR891268    Homo sapiens / ATAC-seq             (p=1.0, cell line p=0.95)
ACCEPT SRR307898    Homo sapiens / RNA-seq              (p=1.0, cell line p=0.96)
ACCEPT SRR3192396   Homo sapiens / RNA-seq              (p=1.0, cell line p=0.93)
ACCEPT SRR1039508   Homo sapiens / RNA-seq              (p=1.0, cell line p=0.75)
ACCEPT SRR5150592   Homo sapiens / RNA-seq              (p=1.0, cell line p=0.76)
ACCEPT SRR031708    Drosophila melanogaster / RNA-seq   (p=1.0, cell line p=0.73)

Rows arrive in whatever order the samples finish, and the probabilities move by a hundredth or so between runs.

Three things worth noticing:

• SRR891268 is labelled ATAC-seq. Its ENA library_strategy field says OTHER — useless — so the assay was recovered from the free-text title alone. That is the case this example exists for: the answer is in the prose, not in the structured field.
• SRR031708 is a fruit fly, worked out from the sample name S2_DRSC_Untreated-1.
• The cell-line answers differ in kind from the assay ones. 0.93–0.96 is a confident yes; 0.73–0.76 is a soft yes worth checking. A yes/no question answers with a probability, so a middling number means genuinely unsure — not "a bit of a cell line".

All six are accepted, and nothing reaches the review queue. These records are clean enough to decide, even with the floor raised to 0.999. The review branch fires when a label's probability falls below params.floor, or when the assay comes back unknown because the metadata never said — which is what you want on a cohort that hasn't been hand-picked.