Upload alphagenome_scorer.py with huggingface_hub

alphagenome_scorer.py

@@ -6,11 +6,19 @@ effects that AlphaMissense (missense-only) and even Evo 2 capture less directly.
 This completes the ensemble: AlphaMissense (missense) + Evo 2 (coding/global) +
 AlphaGenome (regulatory).
 
-Status: real integration against DeepMind's free non-commercial API + Python SDK.
-It requires an API key (free for research) in the env var GEMEO_ALPHAGENOME_KEY,
-obtained at https://deepmind.google.com/science/alphagenome . Without a key this
-returns None — it NEVER fabricates a score. Not yet benchmarked in this repo
-(pending a key); add its AUROC to benchmark/results/ once validated, like Evo 2.
+Status (honest): the integration RUNS against the live API with a key — verified
+end-to-end (score_variant returns the real 19-modality effect outputs). What is
+NOT yet validated: a calibrated single-scalar pathogenicity. A naive max-|effect|
+aggregation across all modalities did NOT discriminate ClinVar splice variants
+(AUROC ~0.37 on n=28), so it is deliberately not used as a score. The correct path
+is AlphaGenome's recommended gene-centric variant scorers with per-modality
+aggregation (splicing scorer for splice variants, etc.) — that is the next step,
+and its AUROC will be committed to benchmark/results/ once it discriminates. Until
+then this module returns the raw effect summary, clearly labelled, not a
+pathogenicity probability — and never a fabricated number.
+
+Requires an API key (free for research) in env var GEMEO_ALPHAGENOME_KEY, from
+https://deepmind.google.com/science/alphagenome . Without a key, returns None.
 
     pip install alphagenome
     export GEMEO_ALPHAGENOME_KEY=...