Batch 31: The BRCA1 Extension — CAPN3/DMD/BRCA1 Variant Classifier

01 — THE AUDIT

Twelve evaluators, three outcomes

Not a binary pass/fail. Two evaluators needed attention — but "needed attention" split into two categories that look identical from the outside and are not the same thing.

Unchanged — needed nothing

Code fix — evaluator logic changed

Data-curation fix — evaluator untouched, fixture data corrected

PVS1

Unchanged

PS1

Unchanged

PM3

Unchanged

PM2

Code fix

PM4

Unchanged

PM5

Unchanged

PS3

Unchanged

BS3

Unchanged

PP3

Unchanged

BP4

Unchanged

BA1

Data-curation fix

BS1

Data-curation fix

9 unchanged

1 code fix (PM2)

2 data-curation fixes (BA1, BS1)

11 / 12 needed zero evaluator code changes

02 — PM2: CODE FIX

The safety net catching something, not just the patch

BRCA1's real spec says PM2 doesn't apply to insertion/deletion variants. The original plan assumed that could be handled by fixture choice alone. It couldn't — and a guard test is what proved it.

Guard test written

test_brca1_fixtures_do_not_exercise_pm4_pm5_or_unenforced_pm2_indel_gaps — asserting no curated BRCA1 fixture accidentally exercises PM4, PM5, or an unenforced PM2-indel call.

Test fails on PM2

PM4/PM5 stayed inert through fixture-shape avoidance alone, as planned. PM2 didn't: two real, cited Ashkenazi-founder frameshift fixtures needed population_evidence anyway, for their own BA1/BS1 handling — so PM2 always ran a real frequency comparison, contradicting the spec's "does not apply" rule.

✓

Gene-gated fix, test passes

One small, additive, opt-in config key — pm2_excludes_indel_delins: true, set only for BRCA1. Checked before any retrieval-status branch. CAPN3/DMD provably unaffected.

Fixture	PM2 result before		PM2 result after
BRCA1_c.68_69delAG	NOT_MET	→	NOT_APPLICABLE
BRCA1_c.5266dup	NOT_MET	→	NOT_APPLICABLE

03 — BA1 / BS1: DATA-CURATION FIX

A different kind of "not for free"

ba1.py and bs1.py have zero BRCA1- or batch-31-specific code — no new branch, no new config key. What needed fixing was which number got curated into the fixture, not the evaluator that reads it.

BS1 threshold — 1 × 10⁻⁴

1.185 × 10⁻⁴raw gnomAD total AF

185delAG, before correction

Below thresholdGrpmax Filtering AF

excludes Ashkenazi Jewish + other small founder groups

WHY IT MATTERED Both evaluators compare overall_af against the gene's threshold before ever consulting ancestry_specific_max_af — so curating only the ancestry field as founder-excluded, the original plan, would not have worked. The raw total alone already crossed BS1's threshold.

"BA1/BS1's founder-frequency handling needed the right value curated into overall_af for BRCA1's two Ashkenazi-founder fixtures, not a code or config change — the evaluators themselves were already correct and untouched."

Audit of all twelve ACMG/AMP evaluators against BRCA1's real specification, showing which needed changes and what kind

How a guard test caught PM2 incorrectly firing on BRCA1 indel variants, and the gene-gated code fix that followed

How BRCA1's founder-population BA1 and BS1 evaluators were kept correct through fixture data curation rather than a code change