DrugAgent

Multi-agent workflow for drug-target interaction (DTI) evidence. It combines ML scores (DeepPurpose), KG signals, and PubMed RAG evidence, then produces a reasoning tree and final label.

Requirements

• Python 3.11+
• uv

Setup (Repo Root)

curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv
uv sync
source .venv/bin/activate

Environment Variables

Create a .env file (see .env.example) or export these variables:

AZURE_OPENAI_API_LLM_KEY=...
AZURE_OPENAI_ENDPOINT=...
AZURE_OPENAI_DEPLOYMENT_NAME=...
AZURE_OPENAI_API_VERSION=...
AZURE_OPENAI_EMBEDDING_DEPLOYMENT_NAME=...
AZURE_OPENAI_EMBEDDING_API_VERSION=...

Embedding variables are required for PubMed RAG.

Run

Full DrugAgent run:

uv run python -m drugagent.cli --drug Imatinib --gene KIT --enabled_agents ML,KG,RAG

Outputs

• output/trees/{config_id}/{drug}_{gene}.json reasoning trees
• output/summary_{ablation}.csv (CLI)
• output/summary.csv (legacy summary output)
• output/ml_dti_scores
• output/ml_lookup_cache
• output/rag_dti_cache.csv
• output/graph_dti_cache.csv

Data Assets

• ML: DeepPurpose model downloads automatically if not present.
• RAG: place files at data/kinase_rag_index.faiss and data/kinase_rag_metadata.json. If missing, the app tries to download via DRUGAGENT_RAG_GDRIVE_URL into DRUGAGENT_RAG_DOWNLOAD_DIR.
• KG: provide a local KG CSV and set DRUGAGENT_KG_PATH or place it at data/KG+BDB.csv.gz.

Citation

@article{inoue2025drugagent,
  title={Drugagent: Multi-agent large language model-based reasoning for drug-target interaction prediction},
  author={Inoue, Yoshitaka and Song, Tianci and Wang, Xinling and Luna, Augustin and Fu, Tianfan},
  journal={ArXiv},
  pages={arXiv--2408},
  year={2025}
}