DrugAgent is a multi-agent LLM framework that unifies ML programming with biomedical expertise to address the demands of modern drug discovery. It integrates two primary agents: (1) an LLM Planner, which manages the high-level generation and refinement of solution ideas, and (2) an LLM Instructor, which translates these ideas into concrete code, drawing on domain-specific knowledge to address the complex needs of drug discovery tasks.
DrugAgent is built upon the MLAgentBench project. We thank the original authors for their valuable work. To install DrugAgent, follow the steps below:
-
Clone the Repository
git clone https://anonymous.4open.science/r/drugagent-5C42.git cd drugagent -
Create a New Conda Environment
conda create --name drugagent python=3.10 conda activate drugagent
-
Install Dependencies
pip install -r requirements.txt
To run our drugagent on admet task with openai API using gpt-4o-mini:
python -u -m drugagent.runner --task admet --device 0 --log-dir first_test --work-dir workspace > log 2>&1 --llm-name openai/gpt-4o-mini --edit-script-llm-name openai/gpt-4o-mini --fast-llm-name openai/gpt-4o-mini
This will produce logs in first_test directory with the following structure
first_test/
agent_log/
Planner_log # log showing Planner agent's research process
Insturctor_log # log showing Instructor agent's research process
agent_*.json # saved agent states
...
env_log/
tool_logs/
traces/ # snap shots of the agent workspace
trace.json # interaction trace of the agent
...
To run evaluation:
python -m MLAgentBench.eval --log-folder <log_folder> --task <task_name> --output-file <output_name>
This will evaluate all runs under <log_folder> as a json.
Note: As this is an early version intended for research use, some code styles may not be fully polished. We will continue to refine and update the project in future releases.