Upload genome_to_disease.py with huggingface_hub

genome_to_disease.py

@@ -18,7 +18,7 @@ import json, os, sys
 sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
 from rare_disease_kg import load_kg
 from pillar_a_kg_proposer import propose
-from genomic_scorer import top_pathogenic_variant, pathogenicity
+from genomic_scorer import top_pathogenic_variant, pathogenicity, VEP_VALIDATION
 
 # cohort causal gene → ground-truth disease (ORPHA)
 GENE_DISEASE = {
@@ -36,15 +36,16 @@ def genome_to_disease(kg, gene, variant_path, k=3):
     return out.get("diseases", [])
 
 
-def _demo(online=True):
+def _demo():
     kg, src = load_kg()
     print("=" * 82)
     print(f"GEMEO — genome → disease   [variant data: real AlphaMissense+ClinVar | KG: {src}]")
     print("=" * 82)
     hits1, n, results = 0, 0, {}
     for gene, truth in GENE_DISEASE.items():
-        v = top_pathogenic_variant(gene, online=online)
-        if not v:
+        v = top_pathogenic_variant(gene)
+        if not v or v.get("alphamissense") is None:
+            print(f"  {gene}: no live variant data (offline?) — skipped")
             continue
         path = v["alphamissense"]
         diseases = genome_to_disease(kg, gene, path, k=3)
@@ -60,16 +61,17 @@ def _demo(online=True):
             print(f"    → genome→disease #1: {top} ({kg.names.get(top, top)})  {flag}")
             print(f"       evidence: {diseases[0]['evidence']}")
 
-    # benign control: a benign variant should NOT score as pathogenic
-    b = top_pathogenic_variant("FBN1_benign", online=False)
-    print(f"\n  [control] FBN1 benign variant {b['variant']}  AlphaMissense={b['alphamissense']:.3f} "
-          f"({b['clinvar']}) → pathogenic? {b['alphamissense'] >= PATHO_THRESH}  (correctly NO)")
+    # discrimination is validated separately on 30 real ClinVar variants:
+    print(f"\n  [validation] variant-effect AUROC on {VEP_VALIDATION['n']} real ClinVar "
+          f"{VEP_VALIDATION['gene']} variants (P vs B): "
+          f"AlphaMissense {VEP_VALIDATION['alphamissense_auroc']} · Evo 2 7b {VEP_VALIDATION['evo2_7b_auroc']} "
+          f"(benchmark/results/evo2_vep_fbn1.json)")
 
     acc = hits1 / n if n else 0.0
     print(f"\n  genome→disease Top-1 recovery: {hits1}/{n} = {acc*100:.0f}%  "
           f"(real pathogenic variant → correct rare disease)")
     res = {"top1_recovery": acc, "n": n, "patho_threshold": PATHO_THRESH,
-           "benign_control_pathogenic": b["alphamissense"] >= PATHO_THRESH, "per_gene": results}
+           "vep_validation": VEP_VALIDATION, "per_gene": results}
     json.dump(res, open("/tmp/genome_to_disease_demo.json", "w"), indent=2)
     print("  Saved /tmp/genome_to_disease_demo.json")
     print("\n  Next (Mayo substrate): condition the world model (Pillar B) on this genomic")
@@ -78,4 +80,4 @@ def _demo(online=True):
 
 
 if __name__ == "__main__":
-    _demo(online="--offline" not in sys.argv)
+    _demo()