DrugAgent

Multi-agent evidence integration and reasoning for drug–target interaction (DTI) prediction

DrugAgent brings together machine-learning predictions, knowledge-graph evidence, and biomedical literature to assess candidate drug–target interactions. Instead of presenting a model score alone, the workflow combines evidence from specialized agents and produces a reasoning tree alongside a final prediction.

Paper (arXiv) · Demo: explore saved results · Quick start · Outputs

How it works

Given a drug and a target gene, DrugAgent can gather complementary evidence through three components:

Component	Evidence source	Role
ML	DeepPurpose	Provides model-based DTI scores.
KG	Local knowledge graph	Provides graph-based evidence about the interaction.
RAG	PubMed literature index	Retrieves relevant biomedical literature evidence.

The workflow organizes the collected signals into a reasoning tree and a final interaction label.

Drug + target gene
       |
       +---- ML agent  ---- DeepPurpose prediction
       |
       +---- KG agent  ---- Knowledge-graph evidence
       |
       +---- RAG agent ---- Retrieved literature evidence
       |
       v
Evidence integration / multi-agent reasoning
       |
       +---- Reasoning tree (JSON)
       +---- Prediction summary (CSV)

The agents are selected with the --enabled_agents option. The full example below enables all three.

Quick start

Requirements

• Python 3.11+
• uv

Install

Run from the repository root:

git clone https://github.com/inoue0426/DrugAgent.git
cd DrugAgent
uv sync

Configure credentials

Copy the example environment file and supply the credentials for the services you intend to use:

cp .env.example .env

The example file documents settings for Azure OpenAI, Claude, and NCBI. In particular, the PubMed RAG workflow requires embedding configuration. Do not commit API keys or a populated .env file.

Run a drug–target example

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

This is an example invocation, not a guaranteed zero-configuration demo: ML model files, KG data, RAG assets, and the relevant service credentials must be available as described below. The repository includes a local KG asset, while the literature index may need to be obtained separately.

Browser demo: saved evidence explorer

A public, read-only version is available on GitHub Pages without credentials or hosting fees. It shows the 11 saved examples and does not run new inference.\n\nDrugAgent includes a local, read-only Gradio viewer for the saved drug–target examples in output/summary_v0_ml_kg_rag_nomodel_none.csv and output/input_payloads/. Choose a drug–target pair to inspect its recorded combined label, explanation, and ML/KG/RAG evidence. No API keys, model downloads, or new inference are needed.

From the repository root:

uv run --no-project --python 3.11 --with gradio==6.5.1 python drugagent_demo.py

Open the local URL printed in the terminal (usually http://127.0.0.1:7860). This is a local browser server, not a public website. The saved examples are displayed as recorded; their labels and confidence are not independently validated here. To run predictions on new pairs, follow the CLI setup above, including its model, data, and credential requirements.

Data and model assets

Component	Required resource	Configuration
ML	DeepPurpose model	Downloaded automatically if missing.
KG	Knowledge-graph CSV	Set DRUGAGENT_KG_PATH, or use data/KG+BDB.csv.gz.
RAG	FAISS index and metadata	Provide data/kinase_rag_index.faiss and data/kinase_rag_metadata.json, or configure the paths in .env.example.

If the RAG index is not present, the application can attempt to download it using DRUGAGENT_RAG_GDRIVE_URL and DRUGAGENT_RAG_DOWNLOAD_DIR. Check .env.example for available overrides.

Outputs

DrugAgent writes predictions, intermediate evidence, and evaluation artifacts under output/. Not every file is produced by every command.

Path	Contents
output/trees/{config_id}/{drug}_{gene}.json	Reasoning tree for a drug–target pair.
output/summary_{ablation}.csv	CLI prediction summary for an agent configuration.
output/summary.csv	Legacy summary output.
output/input_payloads/{drug}__{target}.json	Saved evaluation input payload.
output/ml_dti_scores/	ML prediction artifacts.
output/ml_lookup_cache/	Cached ML lookups.
output/rag_dti_cache.csv	Cached literature evidence.
output/graph_dti_cache.csv	Cached graph evidence.

For an actual example of a saved result, see the sample summary CSV. These files are artifacts of the workflow; they should not be interpreted as a claim of validated DTI performance without the evaluation protocol in the paper.

Plausibility and faithfulness evaluation

The repository also provides a script for evaluating the plausibility and faithfulness of saved reasoning inputs.

From a prediction summary CSV:

uv run python src/faithfulness_plausibility_eval.py \
  --summary-csv output/summary_v0_ml_kg_rag_nomodel_none.csv \
  --output output/plausibility_faithfulness_from_summary.jsonl

The script reads the input_payload column (either JSON content or a path such as output/input_payloads/*.json).

From JSONL inputs:

uv run python src/faithfulness_plausibility_eval.py \
  --input data/plausibility_faithfulness_demo.jsonl \
  --output data/plausibility_faithfulness_results.jsonl

These evaluations use AZURE_OPENAI_API_LLM_KEY and CLAUDE_ENDPOINT from .env; see src/drugagent/kinase/config_utils.py. They are separate from the basic drug–target inference command.

Repository guide

Location	Purpose
src/drugagent/cli.py	Command-line entry point used in the quick start.
src/drugagent/	Packaged agent orchestration and related modules.
src/	Supporting utilities and evaluation scripts.
data/	Bundled reference data and example evaluation inputs.
output/	Stored predictions, evidence, and evaluation results.
.env.example	Configuration variables and optional resource paths.

The repository includes both packaged modules and older top-level scripts. For the documented inference workflow, begin with python -m drugagent.cli.

Paper and citation

For the experimental methodology, datasets, baselines, and reported performance, refer to the DrugAgent paper. The commands above demonstrate inference and reasoning evaluation; they are not a complete end-to-end reproduction script for every paper experiment.

@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}
}

License

Released under the MIT License.