Sorting hat: nine runs with PA14 active, curated-label check, class breakdown and scale

[hyphaltip]

What this adds

Report: docs/reports/2026-10-07-classifier-runs-and-scale.md. Data: docs/reports/data/sorting_hat/classifier_check/ and .../runs/. Scripts: analysis/sorting_hat_run/ (class_breakdown.py, curated_label_check.py).

• Eight existing proteomes re-run on main with PA14 the only active family, plus a new genome (Blastomyces dermatitidis ER-3, 8,107 proteins, no truth set, all statuses unvalidated).
• Class breakdown per proteome (signal peptide, candidate, no-mechanism, not-surface, repeat, PA14 domain, allergen, antigen).
• Exploratory check against curated labels. R0 removes intracellular proteins (1 of 56 C. albicans regulators called) and keeps almost every secreted wall protein (27 of 28 S288C hard negatives called). On proteins that did not tune a module, the repeat call fires for 5 of 27 S288C hard negatives and for 0 of 3 strong-evidence adhesins. PA14 fires for none of the strong-evidence adhesins. The samples are small (3 and 5 proteins), so these are not estimates of sensitivity or specificity.
• A section that restates the overall scale against the original framing (a breakdown of classes per proteome) and lists what remains.

Read before merging

• Descriptive and exploratory. The check is partly circular (detectors built from the same literature) and flags tuned proteins with a name pattern of my own.
• wall_family_domain = not_called means no hit in an active family. Only PA14 is active.
• The BAD1 ortholog in B. dermatitidis is called a candidate by both repeat detectors. BAD1 was a control in earlier detector work, so that is a sanity check, not a test.
• No code in the package changed. No family table change. Run outputs are copies of _workdir results.

Pre-commit (ruff, ruff-format) passes.

🤖 Generated with Claude Code