ProteinSwarm: Swarms of Large Language Model Agents for Protein Sequence Design

Summary

ProteinSwarm introduces a novel approach to protein sequence design by leveraging a swarm of Large Language Model (LLM) agents, each responsible for optimizing a single residue within a target sequence. Through multi-agent collaboration, the system iteratively refines protein sequences to achieve user-defined objectives.

Project Structure

swarm/                  
├── design_loop.py                # Main design iteration engine
├── config.py                     # Configuration and data structures
├── llm_interface.py              # Agent promt
├── memory_system.py              # Memory system
├── structure_utils.py            # Structure analysis utilities
├── folding_utils.py              # Protein folding
├── structure_evaluation.py       # Design objective evaluation
├── rosetta_energy_utils.py       # Energy calculation
├── constants.py                  # Configuration constants
├── requirements.txt              # Python dependencies
├── analysis/                     # Analysis tools
│   ├── energy_monitor.py         # Energy monitor
│   ├── llm_comparison.py         # LLM comparison
│   ├── sequence_logo.R           # Sequence logo plot
│   └── sequence_space.R          # Sequence space plot
├── examples/                     # Example run scripts
│   ├── run_goals.sh              
│   └── test_design_goals.py    
└── logs/                         # Design run logs

Quick Start

1. Installation

# Clone the repository
git clone git@github.com:lamm-mit/ProteinSwarm.git
cd ProteinSwarm

# Create virtual environment
conda create -n ProteinSwarm python=3.10
conda activate ProteinSwarm

# Install core dependencies
pip install -r requirements.txt

Set up OmegaFold for structure prediction.

2. Configuration

Create a local_config.py file with your API credentials:

OPENAI_API_KEY = "your-openai-api-key"

3. One-Click Design

For a quick start, use the automated pipeline:

Run a single goal with custom iterations (default 64):

python test_design_goals_batch.py 32 --goal beta_strands

Available Design Goals

• alpha_helices_hydrophilic - Form alpha helices using hydrophilic residues
• beta_strands - Form beta strands with alternating hydrophobic/polar residues
• loose_coils - Form loose, extended coils with polar/charged residues
• alpha_helices_alanine_leucine_glutamate - Form alpha helices with specific amino acids

Running with Energy Monitoring

Use the bash script to run all goals with real-time energy monitoring:

bash run_all_goals_with_monitoring.sh 64

References

@article{buehler2025swarms,
  title={Swarms of Large Language Model Agents for Protein Sequence Design with Experimental Validation},
  author={Wang, Fiona Y. and Lee, Di Sheng and Kaplan, David L. and Buehler, Markus J.},
  journal={arXiv preprint arXiv:submit/7027477},
  year={2025},
  url={https://arxiv.org/user/},
}