STELLA: Self-Evolving LLM Agent for Biomedical Research

This repository contains the official implementation of the paper STELLA: Self-Evolving LLM Agent for Biomedical Research.

Paper on arXiv | Mention on 𝕏 (Twitter)

---

Abstract

Recent advancements in Large Language Models (LLMs) have demonstrated their potential in specialized fields such as biomedical research. However, their performance is often limited by the lack of domain-specific knowledge and reasoning abilities. To address this, we introduce STELLA, a Self-Evolving LLM Agent designed for complex biomedical research tasks. STELLA leverages a novel self-evolving mechanism that enables it to continuously learn and adapt from new data and experiences. Our agent architecture integrates a dynamic knowledge base, a reasoning module, and a self-correction component, allowing it to perform intricate tasks such as literature analysis, experimental design, and data interpretation. We evaluate STELLA on a comprehensive benchmark of biomedical question-answering and research simulation tasks, demonstrating its superior performance over existing state-of-the-art models. Our results highlight the potential of self-evolving agents to accelerate biomedical discovery and innovation.

Key Figures

Figure 1: Overview of the STELLA framework. The framework consists of four main components: a reasoning template, a manager agent, a dev agent, and a critic agent. The tool ocean provides a set of predefined and self-evolving tools for the agents to use.

Figure 2: Performance of STELLA on various benchmarks. (A) Comparison of STELLA with other LLMs on Humanity's Last Exam (HLE) Biomedicine, LAB-Bench (DBQA), and LAB-Bench (LitQA). (B) Self-evolving performance of STELLA on the same benchmarks with increasing computation budget.

Project Structure

STELLA/
├── 📄 README.md                    # This documentation
├── 🚀 stella_core.py               # Core multi-agent system
├── 🚀 start_stella_web.py          # Web interface launcher
├── 🎨 stella_ui_english.py         # Web UI implementation
├── 💾 memory_manager.py            # Memory management system
├── 📚 Knowledge_base.py            # Knowledge base system
├── 🛠️ predefined_tools.py          # Core system tools
│
├── 📁 new_tools/                   # Professional tool library
│   ├── 📚 literature_tools.py      # Literature search tools
│   ├── 🧬 database_tools.py        # Biomedical databases
│   ├── 💊 virtual_screening_tools.py # Drug discovery tools
│   ├── 🧪 enzyme_tools.py          # Biochemical analysis
│   ├── 🤖 llm.py                   # LLM integration utilities
│   └── 🛡️ biosecurity_tool.py      # Safety mechanisms
│
├── 📁 asset/                       # Images and figures
│   ├── 🌟 stella_logo.png          # STELLA logo
│   ├── 📊 stella_illustration.png   # Framework diagram
│   └── 📈 Stella_result.png        # Performance results
│
└── 📁 prompts/                     # Prompt templates

Usage

Quick Start

Simple mode (basic functionality):

python stella_core.py

Memory-enhanced mode (recommended):

python stella_core.py --use_template --use_mem0

Web interface (user-friendly):

python start_stella_web.py

Usage Options

• Basic Mode: Core STELLA functionality with standard agents
• Memory Enhanced: Adds template learning and enhanced memory capabilities using Mem0
• Web Interface: Provides a user-friendly Gradio interface for interaction

Installation

Prerequisites

# Python 3.8+ required
python --version

# Core dependencies
pip install gradio>=4.0.0
pip install 'smolagents[mcp]'
pip install numpy pandas scikit-learn
pip install requests beautifulsoup4 markdownify

Optional Enhanced Features

# For Mem0 enhanced memory (recommended)
pip install mem0ai

# For biomedical analysis
pip install biopython rdkit-pypi
pip install pymed arxiv scholarly

# For MCP tools integration
pip install uvx

API Keys Configuration

Create a .env file in your local environment with the following API keys:

# Create .env file
touch .env

# Add the following API keys to .env file:
echo "OPENROUTER_API_KEY=your_openrouter_api_key_here" >> .env
echo "SERPAPI_API_KEY=your_serpapi_key_here" >> .env
echo "PAPERQA_API_KEY=your_paperqa_key_here" >> .env

Required API Keys:

• OpenRouter API: For LLM model access and routing
• SERPAPI: For enhanced web search capabilities
• PaperQA API: For academic literature analysis

Get API Keys:

• OpenRouter: https://openrouter.ai/
• SERPAPI: https://serpapi.com/
• PaperQA: https://paperqa.ai/

Resource Download & Setup

Large Resource Files: Due to the size of biomedical datasets, additional resources are available for download:

Download Resources

# Download the resource package (optional for basic usage)
# Manual download from: https://drive.google.com/file/d/1n8I-gkM58wL1OZysdpYhr-Q9ZApd9yU4/view?usp=sharing
# File: resource_backup_20250719_055729.zip (2.0GB compressed)

Extract Resources

# Create resource directory
mkdir -p resource/

# Extract downloaded resources (if downloaded)
unzip resource_backup_20250719_055729.zip -d resource/

# Verify extraction
ls resource/

Note: Resources are optional for basic STELLA functionality but enhance performance for complex biomedical tasks.

Updates & Collaboration

• Wet-Lab Verification: We are currently in the process of conducting wet-lab experiments to verify the key findings and predictions generated by STELLA.
• Call for Collaboration: We welcome both wet-lab collaborations (target discovery, antibody/protein/RNA optimization, etc.) and community contributions including: 🔧 New tools & algorithms 📊 Datasets & knowledge bases 💻 Software integration 📋 Benchmarks & metrics 📚 Tutorials & use cases. Please reach out to zz8680@princeton.edu or submit pull requests!
• Full Release: The complete, full-version preprint and the final source code will be released soon. Stay tuned for updates!

Citation

If you find our work useful, please consider citing our paper:

@article{jin2025stella,
  title={STELLA: Self-Evolving LLM Agent for Biomedical Research},
  author={Jin, Ruofan and Zhang, Zaixi and Wang, Mengdi and Cong, Le},
  journal={arXiv preprint arXiv:2507.02004},
  year={2025}
}