BioMCP

Inspiration

ChatGPT has biological knowledge spawning from its training of all the available biomedical litterature - but if you ask it to combine information from different sources it will give you a very nice plan on what steps to take but will not get the information.

What it does

We implemented a few MCP servers / APIs that provide comprehensive abstractions to access standard biological databases (KEGG, UniProt etc..). and integrated them into Claude desktop to answer a biological research query (starting from the output of a wetlab experiment).

How we built it

mcp server python package / Cursor and claude code.

Challenges we ran into

• Firstly, we wanted to build our own router and planner so that we could do tool calling / MCP server calling ourselves but this did not work out as we could not figure how to consistently get a good plan out of Llma or other together.ai models (the plans always hallucinated tools / added extra unnecssary steps). - - Secondly, generating dymaincally a DAG of steps is pretty challnging

Accomplishments that we're proud of

• Making it work
• beating chat gpt at the task ## What we learned challenges of Agents ## What's next for BioMCP
• integration of biological foundation models to answer target discovery questions.

Built With

• claude
• python