BioDiscoveryAgent is an AI agent for closed loop design of biological experiments. BioDiscoveryAgent designs genetic perturbation experiments using only an LLM (Claude v1) paired with a suite of tools (literature search, gene search, AI critique).

Install required packages using the following command:

pip install -r requirements.txt

Claude API key is required for running the code. Please visit the Anthropic website for more information

1. IFNG
2. IL2
3. Carnevale
4. Scharenberg

To run BioDiscoveryAgent with all tools on the IFNG dataset:

python research_assistant.py  --task perturb-genes-brief --model claude-1 --run_name test --data_name IFNG --steps 5 --num_genes 128 --log_dir v1

To preprocess your own dataset, please follow the instructions in the preprocessing notebook

Please cite our preprint if you use this code in your research:

@article{roohani2024,
  title={BioDiscoveryAgent: An AI Agent for Designing Genetic Perturbation Experiments},
  author={Roohani, Yusuf and Vora, Jian and Huang, Qian and Steinhart, 
  Zachary and Marson, Alexander, Liang, Percy and Leskovec, Jure},
  journal={arXiv preprint},
  year={2024},
}