Upload README.md with huggingface_hub

README.md

@@ -16,9 +16,14 @@ Trained classifier for DPYD variant functional classification (no_function / dec
 - v1: Baseline — 273 training rows, 4 decreased_function examples, F1=0.0 for decreased_function
 - v2: +Offer 2014 + PharmVar data, SMOTE, domain features — 392 rows, 23 decreased_function examples, F1=0.66 (XGB)
 - v3: +AlphaMissense + CADD + VEP annotations on training set — AM importance 0→0.01, sentinel gap closed
+- v4: protein-only model — drop clnsig_norm/activity_pct/AF, keep AM/CADD/SIFT/PolyPhen/domain/aa_position/vep_consequence. 13/25 Scaria variants changed class (no_function→decreased_function). CV accuracy 0.91→0.73 (expected — label proxy removed). Best model: lgbm (decreased_function F1=0.25).
 
 ## Two-tier inference note
-For novel population-discovery variants without ClinVar significance or measured activity (e.g. Scaria 2025 Indian cohort), use AlphaMissense direct score rather than this classifier. The classifier is optimized for variants with clinical labels and measured activity data. See [validation paper](https://doi.org/10.5281/zenodo.20727790).
+The two tier-2 options for novel population-discovery variants (no ClinVar significance, no measured activity — e.g. the Scaria 2025 Indian cohort) are **complementary, not competing**:
+- The **protein-only classifier (v4)** is preferred when class discrimination between `decreased_function` and `no_function` is required. Unlike v1–v3, v4 does not default novel variants to `no_function` (it reclassified 13/25 Scaria variants), at the cost of in-distribution accuracy it does not need on this tail.
+- The **raw AlphaMissense score** is sufficient when the clinical question is simply pathogenicity vs. benign.
+
+For variants that *do* carry clinical labels or measured activity, the v2/v3 mixed-feature classifier (decreased_function F1=0.66) remains the better tool. CPIC-assigned variants are handled by the deterministic engine, not any classifier. See [validation paper](https://doi.org/10.5281/zenodo.20727790).
 
 ## Ground truth sources
 PharmVar activity scores, Offer et al. 2014, GeT-RM 2024 (Gaedigk et al.), ClinVar