[emattei]

crispr_analysis_utils.qc, with its skill. One pass over a channel: find the
floor, measure the cells, cut, write, then look at the guides.

data = md.read_h5mu("results/mudata/mudata.h5mu")
qc.publish("results/qc/knee", *qc.knee(data))
cells, _ = qc.filter_cells(data, min_umis=1000)       # read off that knee
fractions = qc.cell_fractions(cells)
qc.publish("results/qc/metrics", *qc.metric_distributions(fractions))
passes, rules = qc.cell_filters(fractions, nmads=5)
clean = qc.apply_filters(cells, passes, rules, path="results/qc/filtered.h5mu")
qc.publish("results/qc/guides", *qc.guide_abundance(qc.guide_counts(clean)))

Every figure is three files: the PDF at 300 dpi, a tsv.gz of the numbers
it was drawn from, and a caption generated with it so the two cannot drift.
publish refuses a figure with an empty table or a blank caption.

The panels draw the rules the file was cut on. metric_distributions
calls cell_filters for them, so passing both the same nmads, limits and
band makes it impossible for the figure to disagree with the object. Each
line is labelled with its rule and value, and each panel titled with how many
cells it put beyond them -- a cut that removes a handful leaves no grey tall
enough to see, and without the count a one-colour panel looks like a panel
that failed to say anything.

apply_filters writes the record. uns["qc"] carries the floor, every
threshold with its source, and how many cells each rejected. It refuses a
frame that does not cover exactly the object's cells: a cell no rule saw
would otherwise pass every one of them.

Classes come from the contig, not the symbol. MT also names the nuclear
metallothioneins and the MTRNR2L family; on GENCODE v43 the contig gives 37
genes and the prefix 566. Ribosomal pseudogenes are counted apart from the
~80 real ribosomal proteins they share a prefix with.

The guide side reports what a channel saw. The guides never seen get a bar
of their own on the linear strip of a symlog axis, since a log scale has no
place for a zero and leaving them out answers a different question; the
caption says plainly that a channel holds far fewer cells than a library holds
guides, so a guide missing from one is not a guide missing from it.

Plots are Okabe-Ito for categories and cividis for continuous scales, per
CLAUDE.md. mudata_builder gains the uns the QC pass reads.

811 tests.

🤖 Generated with Claude Code