Hydrophobin validation: module split, CFEM/hydrophobin activation, truth set, measurements, relaxed level (not shipped)

[hyphaltip]

Summary

This branch does the hydrophobin work for the sorting hat (cellsurface_sorting_hat) and the changes the tool needs for it: a separate module and calls for hydrophobin and HsbA Pfam families, the CFEM and hydrophobin family sign-offs, a hydrophobin truth set, a measurement of the strict hydrophobin call in 8 proteomes, and a tested relaxed (extended) level that is built but not yet part of any call. It also carries the planning documents and independent reviews. 42 commits, 194 files, +32,209/-30 lines (most of the lines are analysis tables, sequences and documents; src/ is about 300 lines changed plus 2 new modules).

It supersedes PR #75 (CFEM and hydrophobin activation, closed unmerged: #75 put the hydrophobin rows in pfam_adhesion, so hydrophobin hits set wall_family_domain). Its note fix 1466f90 is carried here as 97f2676. per-call-status (#76) is merged to main and this branch contains it.

Behaviour changes to review (src and data)

Change	Files	Effect
Hydrophobin-class Pfam models move to their own module pfam_hydrophobin (PF01185, PF06766, PF22354, PF28987, PF29785, PF29802, PF29465). HsbA (PF12296) goes to pfam_hsba.	data/sorting_hat/family_table.tsv, modules/pfam.py (MODULES), modules/cli.py	A hydrophobin hit no longer sets wall_family_domain or cell_wall_adhesion_candidate.
Two new evidence calls hydrophobin_domain and hsba_domain, both added to the mechanism lists of other_not_surface and other_surface_no_mechanism	categories.yaml	A surface protein with a hydrophobin or HsbA hit is not "no mechanism". config_sha256 changes.
Active Pfam families: PA14 (signed off 2026-10-07), CFEM, Hydrophobin and the other hydrophobin-class models, Hydrophobin_like, HsbA (owner decisions 2026-10-08). Cerato-platanin was dropped as out of scope. Candidate families (Flocculin PF00624, Flocculin_t3 PF13928, Hyr1 PF15789, PIR1-like_C PF22799, ALS_M PF30910) were added inactive with notes.	family_table.tsv (24 families, 10 active)	Activation is evidence only. Call status stays unvalidated unless a status file exists.
Per-module Pfam artefact digest (a change to one Pfam module's rows no longer changes the identity of the others)	modules/cli.py _family_digest(families, module)	The identities of pfam_adhesion and pfam_allergen change once, so any status file tied to the old identities needs re-measuring.
New module cys8_pattern (eight-cysteine spacing from published sets plus the R0 signal peptide)	modules/cys8.py, data/sorting_hat/cys8_spacing.yaml (85 gap entries with source quotes)	Reported side column. Not read by any call.
New module hydrophobin_relaxed and the frozen model file hydrophobin_relaxed.hmm (seven Pfam models, GA 2.60/-1000, --nobias) with provenance	modules/hydrophobin_relaxed.py, data/sorting_hat/hydrophobin_relaxed.{hmm,provenance.json}	The module can be run and is tested. No call reads it, and no job is added to submit_modules.sh. See "What this PR does not do".
Report text for limit 4 now says hydrophobin and HsbA are separate evidence calls	outputs.py, docs/superpowers/specs/2026-10-04-orchestrator-design.md (kept equal, a test checks it)	Text only.
pfam.parse_domtblout also returns seq_score and query	modules/pfam.py	Additive.

Golden files (tests/cellsurface_sorting_hat/golden/) were regenerated after reading every changed line. The only differences are the new calls and modules, the new config_sha256, and the new basis strings.

Analysis and data (analysis/hydrophobin_truth/)

• Truth set: 174 Swiss-Prot entries named hydrophobin or rodlet (UniProt 2026_03): 131 T2 (experimental PubMed evidence on a FUNCTION, SUBCELLULAR LOCATION or SUBUNIT comment), 43 T3, no T1. Pfam is not used for any label. 30 MMseqs2 clusters (28 with a T2 member). Literature lists (Jensen 2010, Xu 2021) were read; none gives protein-level evidence, so they form a held-out "literature, predicted" tier (50 Jensen proteins resolved to sequences).
• Calibration inputs: truth mapped to 16 proteomes (12 used), assumed-negative tables, a stored development/test split, hard-negative sets (CFEM, cerato-platanin, HsbA, PIR/Ccw12, small secreted cysteine-rich; counts in hard_negative_queries.md), R0 SignalP on 1,608 reference sequences.
• Runs: SignalP, TMHMM, Pfam and the new modules on 12 proteomes (3 A. fumigatus strains, C. immitis, B. dermatitidis, C. albicans, S288C, B. bassiana, F. fulva, F. graminearum, P. expansum, P. ostreatus PC9). Run drivers are in analysis/hydrophobin_truth/run/.
• Freeze: freeze.json (commit f1986e2) pre-registers the relaxed level (--nobias, 2.6 bits, R0 called, at least 8 cysteines), the cutoff rule, the per-fold HMM cutoffs, the thresholds and what the author had seen. Only tuning data was scored before it. Large score and HMM files stay untracked; their hashes are in freeze.json.
• Evidence sheets and curation: evidence_sheets.tsv (61 unlabelled calls) and curation_rubric.md (written before any decision).
• Tests: tests/hydrophobin_truth/ (about 90 tests) cover tiers, folds, cutoff rule, derived HMM file vs --cut_ga, alignment checks, evaluation, cost and freeze checks.

Measured results (all smoke; negatives are assumed; leakage partial or tuned on truth)

Result	Value	Report
Strict pfam_hydrophobin (module status) in 8 proteomes	sensitivity 1.00 in 5, 0.80, 0.60 and 0.25 in the others; 3 to 7 positives per proteome; lower bounds 0.00 to 0.57; specificity not informative (19 of 22 unlabelled calls have 8 or more cysteines)	docs/reports/2026-10-08-hydrophobin-validation.md
Published-spacing rescue (cys8_pattern)	recovers 1 of 6 Pfam-missed proteins in the proteomes; matches 78 of 131 T2 (60%)	same report
Relaxed level, leave-cluster-out	recovers 6 of 6 Pfam-missed clusters (Wilson [0.61, 1.00]); extended recall on T2 130 of 131 vs 122 strict	docs/reports/2026-10-08-hydrophobin-extended-level-L5.md
Custom HMM	not testable on the primary set (relaxed leaves 0 clusters unrecovered); leave-cluster-out 58% of T2 with default filters, 90% with --nobias; not part of v1	L5 report
Aligner comparison (mafft L-INS-i, famsa, muscle 5.1)	equivalent within 1 or 2 of 131 proteins; --nobias is what matters	L5 report
Cost of the relaxed level on the 7 test proteomes (limit 5 unlabelled extra calls per 10,000)	failed in 2 of 7: B. bassiana 10.55 and F. graminearum 10.72; the other 5 are within 1.0 to 1.9	docs/reports/2026-10-08-hydrophobin-relaxed-cost-and-sheets.md
Repeat call re-measured after the config change	unchanged: S288C 0.435 [0.067, 0.667] / 0.987; C. albicans 0.462 [0.000, 0.788] / 0.976	ledger row H4 in docs/paper/02

What this PR does not do

• hydrophobin_extended is not added to categories.yaml. By the ship rule frozen in freeze.json, the relaxed level is not added while any condition fails, and the cost condition failed in 2 of 7 test proteomes. Precision also failed (see below). Recall and hard negatives passed.
• The custom HMM is not shipped.
• No status file is committed for a call that does not exist yet.
• Nothing about SOWgp-like proteins (see docs/ROADMAP-2026-10-08-DRAFT.md, milestone M5).

Curation of the 61 unlabelled calls and the ship decision (added 2026-10-08)

analysis/hydrophobin_truth/curation_rubric.md (committed first, a9b7a32), evidence in curation/ (24 proteins with the doublet architecture, 21 strict Pfam, 0 named hydrophobin by a paper or curated record under the mapping rule, 6 reciprocal best hits to known hydrophobins, 8 with evidence of another function; 21 papers read, 98 paper gene IDs mapped by sequence), decisions in curator_decisions.tsv, rule in apply_rubric.py (7 tests), ship rule in ship_rule.py (3 tests) and ship_decision.json. Report: docs/reports/2026-10-08-hydrophobin-curation-and-ship-decision.md.

Group	hydrophobin	not hydrophobin	unresolved
strict and relaxed, unlabelled (18)	16	0	2
strict only, unlabelled (3)	0	0	3
relaxed only, unlabelled (40)	0	8	32

Ship rule: recall pass, hard negatives pass, cost fail (2 of 7 test proteomes), precision fail (0 hydrophobin of 7 resolved clusters among 36 clusters of the relaxed-only calls; 29 unresolved, which means no evidence found, not false positive). The relaxed level is not shipped; the owner confirmed on 2026-10-08: keep the strict call only. A bug found and fixed on the way: the evidence sheet's BLAST column against the known hydrophobins was empty for all rows (a repeated BLAST output field name dropped the column); 24 of 61 rows have a hit.

Caveats

• Every status is smoke. No species has 20 independent positive clusters. The Pfam-missed set is 9 proteins in 6 clusters.
• Specificity rests on assumed negatives. The T2 labels were only keyword-checked.
• Pre-freeze changes (aligner, cysteine-column check) were made before any held-out score existed and are recorded in freeze.json.
• The owner could not curate the 61 evidence sheets, so they were curated from evidence and literature (tier T4, "not owner-reviewed") under a rubric committed before any decision. See the new section below. The owner can audit and overrule any row.

Documents (planning and review trail)

Specs (docs/superpowers/specs/2026-10-08-hydrophobin-*.md): validation design rev 3 and custom-HMM/extended-level design rev 4, each with independent reviews. Plans (docs/superpowers/plans/2026-10-08-hydrophobin-*.md) with plan reviews. Reports in docs/reports/. Paper notes updated: docs/paper/02 (ledger rows H1 to H4), 04, 05 (literature, prior art, cysteine spacing quotes). A hand-off: docs/HANDOFF-2026-10-08-hydrophobin.md. A whole-project roadmap draft: docs/ROADMAP-2026-10-08-DRAFT.md.

How to review

1. src/ and data/ first (about 10 files): categories.yaml, family_table.tsv, modules/hydrophobin_relaxed.py, modules/cys8.py, modules/cli.py, pfam.py, and the golden diffs.
2. Check the tests: PYTHONPATH=$PWD/src python3.12 -m pytest tests/cellsurface_sorting_hat tests/hydrophobin_truth tests/calibration_truth -q gives 910 passed and 9 skipped (the skips are tool tests that need hmmer, mafft or MMseqs2 on PATH, and shellcheck). With module load hmmer/3.4 mafft/7.505 mmseqs2/17-b804f the tool tests run too.
3. The numbers in the reports come from analysis/hydrophobin_truth/ scripts. The freeze, evaluate.py, cost.py and the L5 report are the main audit trail.