Skip to main content

View on GitHub

examples/05_drug_target_provenance/
MEDEA is an AI scientist that identifies drug targets from multi-omics data. Two forks run on different diseases (RA and SLE, CD4+ T cells). Mareforma records whether MEDEA’s data pipeline actually ran or whether the answer came from LLM prior knowledge.

​
Recording origin at assertion time

result = run_medea_fork(disease="rheumatoid arthritis", cell_type="CD4")
classification = "ANALYTICAL" if result["generated_code"] else "INFERRED"

graph.assert_claim(
    result["final_hypothesis"],
    classification=classification,
    generated_by="medea/gpt-4o/ra_cd4",
    source_name="medeadb",
)

​
What we found

Both forks returned generated_code = null. Both findings were INFERRED. The classification surfaced a silent pipeline failure immediately, before anyone acted on the results. Led to a bug report: mims-harvard/Medea#6.

​
Setup

python 05_drug_target_provenance.py --install   # ~21 GB MedeaDB download
python 05_drug_target_provenance.py --data
cp Medea/env_template.txt .env                  # set OPENAI_API_KEY
python 05_drug_target_provenance.py --run