BioMedAgent：Autonomous Biomedical data Analysis by LLM-powered multi-agents with self-evolving capabilities

News

• 2026-06-16: The BioMedAgent official website is now available at drai.cn.
• 2026: Our BioMedAgent paper has been published in Nature Biomedical Engineering.

Background

The biomedical field is experiencing a significant expansion in data growth as novel technologies develop, propelled by the vast amounts of medical texts, images, and omics data being collected. Analyzing this data requires complex computational methods that integrates skills from bioinformatics, artificial intelligence, software programming, statistics and mathematics. Various tools and platforms, such as Galaxy and Seven Bridges, provide interfaces for workflow creation and execution, but they face challenges, including reliance on predefined workflows, insufficient support for natural language instructions and lack of summarized readable reports, which restricts intuitive user engagement. Large Language Models (LLMs) have show great advantage in natural language comprehension and demonstrated potential in biomedical data analyses, yet they encounter significant challenges with complex problem planning and low success rate of execution.

Addressing these challenges, we introduce BioMedAgent, an autonomous framework that capitalizes on LLM-powered multi-agents with self-evolving capabilities, incorporating Interactive Exploration (IE) and Memory Retrieval (MR) algorithms. BioMedAgent supports natural language-based task initiation, allowing direct participation from biomedical professionals with no specialized computational or bioinformatics training.

BioMedAgent was tested using the BioMed-AQA bench, comprising 327 manually curated bioinformatic analyses questions across five categories: omics analyses (O), precision medicine support analyses (P), machine learning (M), statistical analyses (S), and data visualization (V). Through the collaborative planning, coding, and execution of multiple agents, it achieved an average success rate of 77%, outperforming two OpenAI web agent applications and a local agent application, each also based on the GPT-4omini LLM. Remarkably, it even surpassed the performance of ChatGPT-4o application using the bigger model, which had a 47% success rate.

Highlights

We propose BioMedAgent, an autonomous biomedical data analyses framework that firstly utilizes LLM-powered multi-agents with self-evolving capabilities.

• We built the first comprehensive benchmark for evaluating LLM applications in biomedical analyses, BioMed-AQA, spanning five categories of tasks, with natural language questions as analyses instructions and manually annotated answers.
• Extensive experiments demonstrate that BioMedAgent outperforms existing OpenAI agent applications, demonstrating higher success rates across diverse biomedical analyses tasks.
• BioMedAgent supports natural language-based task initiation, allowing direct participation from biomedical professionals with no specialized computational or bioinformatics training.

Quickstart

1. Clone this repo

git clone https://github.com/BOBQWERA/BioMedAgent.git

2. Set up environment

Install python

conda create -n BioMedAgent python=3.10
conda activate BioMedAgent

Enabling redis（Assuming that redis is already installed on the system）

redis-server

3. Install python requirements

cd BioMedAgent
python -m pip install -r requirements.txt

4. Set environment variables

export OPENAI_API_KEY="your_openai_api_key"

5. Run demo.py

You can specify the task type via the task parameter. Demo supports four different task instances of ['machine_learning', 'statistics_t_test', 'statistics_qq_plot', 'visualization_survival_plot', 'visualization_violin_plot', 'omics'], which can be called via python demo.py --task task_type

python demo.py --task statistics

For omics, you need to deploy the supporting tool cel2matrix environment for BioMedAgent call, the tool code and documentation information are provided in the tool folder, the code mentioned docker image biogpt_r can be downloaded through Baidu Drive

For statistics_t_test and statistics_t_test, you need to deploy the supporting tool survival_curve and t_test environment for BioMedAgent call, the tool code and documentation information are provided in the tool folder, the code mentioned docker image bio_r can be downloaded through Baidu Drive

Enrich your tools

You are free to extend the collection of tools to enhance the boundary capabilities of BioMedAgent. Refer to the cel2matrix, survival_curve and t_test tool information in the tool folder, you can configure your local environment to add any tools you want. Simply complete the documentation as well as the code, and BioMedAgent will automatically perceive the new tool and use it appropriately in new question. For the BioMedAgent configuration in our paper, its full tool information can be seen in tool_info.json.

Updates

• 2026: Our BioMedAgent paper has been published in Nature Biomedical Engineering.
• 2025-02-27: Add the demo of machine learning, statistics, visualization and omics.
• 2025-02-27: Add the demo of survival plot and t-test.

Citation

If you find BioMedAgent useful in your research, please cite:

@article{bu2026empowering,
  title={Empowering AI data scientists using a multi-agent LLM framework with self-evolving capabilities for autonomous, tool-aware biomedical data analyses},
  author={Bu, Dechao and Sun, Jingbo and Li, Kun and He, Zihao and Huang, Wei and Hu, Jinlin and Zhang, Shanshan and Lei, Shuangshuang and Huo, Peipei and Wang, Zhihao and others},
  journal={Nature Biomedical Engineering},
  pages={1--16},
  year={2026},
  publisher={Nature Publishing Group UK London}
}