atomica-mcp

🏆 HackAging.ai Submission
Longevity Genie Team submission to the Sequence to Function Track at HackAging.ai

📺 Watch Demo Video | 📊 Presentation Slides

What we built:

• This repository: Model Context Protocol (MCP) server - a standardized way for AI assistants and agents to access specialized data and tools. Works with Claude Desktop, Cursor, Windsurf, VS Code with Copilot, and other MCP-compatible AI systems.
• Dataset: 94 analyzed longevity protein structures on Hugging Face
• Custom ATOMICA fork: Polished the original model with improved usability, additional features, and longevity-focused applications
• Analysis examples: NRF2/KEAP1 structural analysis demo (showcased in demo video) with figures showing disease mutations, glycation sites, and critical residues

The Problem

Protein engineering for aging interventions requires synthesizing sequence-to-function knowledge from diverse sources—structural data, interaction predictions, evolutionary conservation, and experimental outcomes. AI agents need standardized access to these resources to generate comprehensive sequence-to-function analyses at scale.

Traditional computational methods are too slow for rapid iteration, and results aren't easily accessible within AI-assisted research workflows.

Our Solution

We built ATOMICA-based infrastructure enabling AI assistants to rapidly screen protein interactions and identify critical functional residues for automated sequence-to-function analysis.

Core Deliverables (Hackathon Focus)

1. atomica-mcp Server (this repository)

• MCP interface connecting AI assistants to interaction scores and critical residue predictions
• Instant access to precomputed ATOMICA analyses for structural insights
• PyMOL visualization scripts for structural analysis
• PDB structure search and metadata resolution (UniProt, Ensembl, gene symbols, organisms)

2. Longevity Proteins Dataset

• Curated structures on HuggingFace with pre-computed ATOMICA analyses
• Residue rankings and critical interaction scores
• Visualization scripts ready for AI-assisted exploration

3. Enhanced ATOMICA Fork

• Polished model with improved usability and additional features
• Longevity-focused applications for high-throughput screening
• Applied to aging-related protein families

Broader Integration

The atomica-mcp server works alongside our biological MCP ecosystem:

• Our servers: synergy-age-mcp, biothings-mcp, opengenes-mcp
• Community servers: omnipath-next, BioContextAI Knowledgebase

This enables AI agents to combine rapid ATOMICA interaction screening with pathway databases, variant annotations, and evolutionary data—synthesizing comprehensive articles about protein engineering opportunities for longevity.

Note: Servers are registered in BioContextAI Registry. Our team are co-authors of the BioContextAI Registry paper (accepted to Nature Biotechnology)

Demo Analysis Results

Our NRF2/KEAP1 analysis (shown in demo video) demonstrates the workflow:

• Confirmed E82 in ETGE motif as the most critical residue for KEAP1 binding
• Found F468 maintains structure near age-affected glycation sites
• Showed R499 glycation has more functional impact than K462
• Identified potentially novel regulatory cysteines C406/C368 in KEAP1

Key advantage: ATOMICA provides fast interaction evaluation to prioritize functionally important residues before applying more computationally intensive structural methods.

---

About This Server

This is a Model Context Protocol (MCP) server providing access to the ATOMICA longevity proteins dataset and PDB structure analysis tools.

ATOMICA is a geometric deep learning model trained on 2M+ interaction complexes. It identifies critical residues in protein structures by analyzing atomic-scale interaction patterns.

Dataset: 94 structures across 5 longevity-related protein families (NRF2, KEAP1, SOX2, APOE, OCT4)
Repository: longevity-genie/atomica_longevity_proteins

Each structure includes:

• Structure file (CIF format)
• ATOMICA interaction scores
• Critical residue rankings
• PyMOL visualization commands

Quick Start

Installation

# Install uv if needed
curl -LsSf https://astral.sh/uv/install.sh | sh

# Run the server
uvx atomica-mcp

Configure with Your AI Assistant

Add to your MCP config file (Claude Desktop, Cursor, Windsurf, etc.):

Cursor/Windsurf: ~/.cursor/mcp.json or ~/.windsurf/mcp.json
Claude Desktop (macOS): ~/Library/Application Support/Claude/claude_desktop_config.json
Claude Desktop (Windows): %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "atomica-mcp": {
      "command": "uvx",
      "args": ["atomica-mcp@latest"],
      "env": {
        "MCP_TRANSPORT": "stdio",
        "MCP_TIMEOUT": "600"
      }
    }
  }
}

Restart your AI assistant. The dataset (~500MB) downloads automatically on first use.

Example Queries

Once connected, try asking your AI assistant:

• "What structures are available for KEAP1?"
• "Show me critical residues in NRF2 ETGE motif"
• "Find APOE structures and their variants"
• "Get PyMOL commands for structure 2FLU"

Available Tools

The server provides 8 MCP tools:

Dataset queries:

• atomica_list_structures - List all structures
• atomica_get_structure - Get details for a specific PDB
• atomica_search_by_gene - Search by gene symbol (e.g., "KEAP1")
• atomica_search_by_uniprot - Search by UniProt ID
• atomica_search_by_organism - Filter by organism

Auxiliary tools:

• atomica_resolve_pdb - Get metadata for any PDB ID
• atomica_get_structures_for_uniprot - Get all PDB structures for a UniProt ID
• atomica_dataset_info - Dataset statistics

For detailed tool documentation, see the full documentation below.

Dataset Management CLI

The package includes a CLI for managing the ATOMICA dataset:

Download Dataset

# Download full dataset
dataset download

# Download to custom directory
dataset download --output-dir data/inputs

# Download only CIF structure files
dataset download --pattern "*.cif"

# Download only files for specific PDB (e.g., 6ht5)
dataset download --pattern "6ht5*"

# Force re-download even if files exist
dataset download --force

List Available Files

# List all files in the dataset
dataset list-files

# Filter by pattern
dataset list-files --pattern "*.cif"

Create/Update Index

# Create index with basic file paths
dataset index

# Create index with full metadata resolution
dataset index --include-metadata

# Custom paths
dataset index --dataset-dir data/atomica --output data/index.parquet

Reorganize Dataset

# Reorganize files into per-PDB folders
dataset reorganize

# Dry run to see what would be done
dataset reorganize --dry-run

Dataset Information

# Show dataset information
dataset info

Development

Python Library Usage

from atomica_mcp.server import AtomicaMCP
from atomica_mcp.mining.pdb_metadata import get_structures_for_uniprot

# Initialize server
mcp = AtomicaMCP()

# Query dataset
structures = mcp.list_structures(limit=10)
keap1 = mcp.search_by_gene("KEAP1")

# Get structures for UniProt
p53 = get_structures_for_uniprot("P04637")

Testing

uv run pytest

Requirements

• Python 3.11+
• All dependencies managed by uv - just run uv sync

Related Projects

• opengenes-mcp - Aging and longevity genetics database
• gget-mcp - Genomics and sequence analysis toolkit
• holy-bio-mcp - Unified bioinformatics framework

License

MIT License - see LICENSE file for details

Contributors

This project was developed by the Longevity Genie Team for the HackAging.ai hackathon:

Anton Kulaga

• ATOMICA fork development
• ATOMICA MCP server
• Presentation
• Hackathon submission

Newton Winter

• Demo analysis (NRF2/KEAP1)
• PDB structure processing
• ATOMICA fork DevOps
• HuggingFace dataset preparation

Livia Zaharia

• Video editing
• Presentation design

Citation

If you use atomica-mcp in your research, please cite:

@software{atomica-mcp,
  title={atomica-mcp: MCP server for ATOMICA longevity proteins dataset},
  author={Kulaga, Anton and Winter, Newton and Zaharia, Livia},
  year={2025},
  url={https://github.com/longevity-genie/atomica-mcp},
  note={HackAging.ai Submission - Longevity Genie Team}
}