Research & personal
Biomedical knowledge graph + AI agent
Independent research project: measuring whether an AI agent answers biomedical questions better when it can query a knowledge graph.
Context
Does an AI agent answer biomedical questions better when it can query a knowledge graph? That is the question behind this independent research project. The disease studied is rheumatoid arthritis.
What I did
- Imported the PrimeKG biomedical knowledge graph into Neo4j.
- Exploration and Cypher queries in Jupyter notebooks.
- Integrated JeuxDeMots, LIRMM’s lexical network.
- Method: steering Claude Code through documented phases (rules, checklists, prompts) while keeping scientific decisions in my own hands.
Results
Research in progress. So far:
- A reproducible Python pipeline populates a Neo4j graph of about 4 million relationships.
- An evaluation design measures whether graph-grounded retrieval (an approach related to RAG) improves the accuracy of the agent’s answers.
- A separate data-verification step checks the graph’s quality before evaluation.