variant-confidence v0.1.0
Calibrated confidence layer for protein variant-effect pathogenicity predictions (AlphaMissense / ESM-1v / EVE). Does not train a new model — adds an auditable calibration layer on top of existing predictors.
Verified in clean clone (no network)
ruff check .→ All checks passed!pytest tests/→ 28 passed in 8.90s
What's in 0.1.0
- Calibration (AC1, AC1b): Platt scaling, isotonic regression, and split/Mondrian conformal prediction.
- ECE metric (AC2, AC9): equal-width + adaptive bins, bootstrap CI, low-count bins flagged as low-reliability.
- Leakage-free split (AC3): temporal by ClinVar release date + gene isolation; index-alignment bug fixed (silent-liar class).
- Missing data (AC4, T13b): explicit missing-score handling + structured-missing detection.
- Non-deceptive reporting (AC7): CLI emits score + interval + method + ECE before/after.
- AlphaMissense join (T13): license-safe integration.
- Degenerate-ECE bug fixed: synthetic generator is now discriminative; AUC preserved after calibration (no collapse to base rate).
License
AGPL-3.0-or-later — Pedro Sordo Martínez (amurlaniakea@gmail.com), 2026.
Install
pip install variant-confidence
variant-confidence --help