For each of 6 diseases, the Open Targets GraphQL API is queried for gene-disease associations. Genes with genetic_association or genetic_literature evidence AND an approved drug (maxClinicalStage == APPROVAL) are selected. The top 2–3 per disease by genetics evidence score form the gold standard ground truth. Nothing is hardcoded — the gold standard is fully discovered at runtime.

GWAS Catalog is queried for genome-wide significant associations (p < 5×10⁻⁸) per disease. Gene symbols are extracted from authorReportedGenes in each locus. Any gene already in the gold standard is excluded. Up to 30 orphan genes per disease (no approved drug, but GWAS signal) are retained.

Each orphan gene is enriched with:

• open_targets_score — max association score across top 3 diseases
• druggability_score — SM tractability tier from Open Targets
• pubmed_count_5yr — PubMed publication count (last 5 years)
• eqtl_effect — proxy: (gwas_p/50)×0.5 + ot_score×0.5
• mr_z_score — proxy: (gwas_p/10)×(1+n_studies/20)
• tissue_specificity — disease-area lookup
• ppi_degree — binned from PubMed count

agent.py runs 10 pre-defined experiments. Each experiment rewrites scorer.py with new feature weights and interaction terms, evaluates recall@K against the discovered gold standard, and keeps the change only if recall improves. Results and scorer snapshots are committed to GitHub after every iteration.