MODEL_CARD.md

metadata

language:
  - en
license: gpl-3.0
tags:
  - molecular-docking
  - drug-discovery
  - distributed-computing
  - autodock
  - boinc
  - computational-chemistry
  - bioinformatics
  - gpu-acceleration
  - distributed-network
  - decentralized
datasets:
  - protein-data-bank
  - pubchem
  - chembl
metrics:
  - binding-energy
  - rmsd
  - computation-time
library_name: docking-at-home
pipeline_tag: boinc

Docking@HOME: Distributed Molecular Docking Platform

Model Card Authors

This model card is authored by:

• OpenPeer AI - AI/ML Integration & Cloud Agents Development
• Riemann Computing Inc. - Distributed Computing Architecture & System Design
• Bleunomics - Bioinformatics & Drug Discovery Expertise
• Andrew Magdy Kamal - Project Lead & System Integration

Model Overview

Docking@HOME is a state-of-the-art distributed computing platform for molecular docking simulations that combines multiple cutting-edge technologies to democratize computational drug discovery. The platform leverages volunteer computing (BOINC), GPU acceleration (CUDPP), decentralized networking (Distributed Network Settings), and AI-driven orchestration (Cloud Agents) to enable large-scale molecular docking at unprecedented speeds.

Key Features

• 🧬 AutoDock Integration: Industry-standard molecular docking engine (v4.2.6)
• 🚀 GPU Acceleration: CUDA/CUDPP-powered parallel processing
• 🌐 Distributed Computing: BOINC framework for global volunteer computing
• 🔗 Decentralized Coordination: Distributed Network Settings-based task distribution
• 🤖 AI Orchestration: Cloud Agents for intelligent resource allocation
• 📊 Scalable: From single workstation to thousands of nodes
• 🔒 Transparent: All computations recorded on distributed network
• 🆓 Open Source: GPL-3.0 licensed

Architecture

Docking@HOME employs a multi-layered architecture:

1. Task Submission Layer: Users submit docking jobs via CLI, API, or web interface
2. AI Orchestration Layer: Cloud Agents optimize task distribution
3. Decentralized Coordination Layer: Distributed Network Settings ensure transparent task allocation
4. Distribution Layer: BOINC manages volunteer computing resources
5. Computation Layer: AutoDock performs docking with GPU acceleration
6. Results Aggregation Layer: Collect, validate, and store results

Intended Use

Primary Use Cases

• Drug Discovery: Virtual screening of compound libraries against protein targets
• Academic Research: Computational chemistry and structural biology studies
• Pandemic Response: Rapid screening for therapeutic candidates
• Educational: Teaching molecular docking and distributed computing concepts
• Benchmark: Testing distributed computing frameworks and GPU performance

Out-of-Scope Use Cases

• Clinical diagnosis or treatment recommendations
• Production pharmaceutical manufacturing decisions without expert validation
• Real-time emergency medical applications
• Replacement for experimental validation

Technical Specifications

Input Format

• Ligands: PDBQT format (prepared small molecules)
• Receptors: PDBQT format (prepared protein structures)
• Parameters: JSON configuration files

Output Format

• Binding Poses: PDBQT format with 3D coordinates
• Energies: Binding energy (kcal/mol), intermolecular, internal, torsional
• Ranking: Clustered by RMSD with energy-based ranking
• Metadata: Computation time, node info, validation hash

Performance Metrics

Benchmark Results (RTX 3090 GPU)

Metric	Value
Docking Runs per Hour	~2,000
Average Time per Run	~1.8 seconds
GPU Speedup vs CPU	~20x
Memory Usage	~4GB GPU RAM
Power Efficiency	~100 runs/kWh

Distributed Performance (1000 nodes)

Metric	Value
Total Throughput	100,000+ runs/hour
Task Overhead	<5%
Network Latency	<100ms average
Fault Tolerance	99.9% uptime

Training Details

This is not a traditional machine learning model but a computational platform. The platform uses:

• AutoDock: Physics-based scoring function (empirically parameterized)
• Genetic Algorithm: For conformational search
• Cloud Agents: Pre-trained AI models for resource optimization

Validation & Testing

Validation Protocol

1. Redocking Tests: Reproduce known crystal structure binding poses (RMSD < 2Å)
2. Cross-Docking: Test on different conformations of same protein
3. Enrichment Tests: Ability to identify known binders from decoys
4. Benchmark Sets: Validated against CASF, DUD-E, and other standard sets

Success Criteria

• RMSD < 2.0 Å: 85% success rate on redocking tests
• Energy Correlation: R² > 0.7 with experimental binding affinities
• Enrichment Factor: >10 for known actives vs decoys
• Reproducibility: 99.9% identical results across multiple runs

Limitations & Biases

Known Limitations

1. Flexibility: Limited receptor flexibility (rigid docking primarily)
2. Solvation: Simplified water models may miss key interactions
3. Metals: Limited handling of metal coordination
4. Entropy: Approximated entropy calculations
5. Post-Dock: Requires expert analysis and experimental validation

Potential Biases

1. Parameter Bias: Scoring function optimized on specific protein families
2. Dataset Bias: Training on predominantly drug-like molecules
3. Structural Bias: Better performance on well-defined binding pockets
4. Resource Bias: GPU access required for optimal performance

Mitigation Strategies

• Provide multiple scoring functions
• Support custom parameter sets
• Enable CPU-only mode for accessibility
• Comprehensive documentation on limitations
• Encourage ensemble docking approaches

Ethical Considerations

Responsible Use

• Open Science: All results timestamped on distributed network for reproducibility
• Attribution: Volunteer contributors credited in publications
• Data Privacy: No personal data collected from volunteers
• Environmental: GPU efficiency optimizations reduce carbon footprint
• Accessibility: Free for academic and non-profit research

Potential Risks

• Dual Use: Could be used for harmful compound design (mitigated by access controls)
• Over-reliance: Results must be validated experimentally
• Resource Inequality: GPU requirements may limit access (mitigated by distributed model)

Carbon Footprint

Estimated CO₂ Emissions

• Single GPU (24h operation): ~5 kg CO₂
• Distributed Network (1000 nodes, 1 year): ~43,800 kg CO₂
• Offset Programs: Partner with carbon offset initiatives
• Efficiency: 20x more efficient than CPU-only approaches

Getting Started

Installation

# Clone repository
git clone https://huggingface.co/OpenPeerAI/DockingAtHOME
cd DockingAtHOME

# Install dependencies
pip install -r requirements.txt
npm install

# Build C++/CUDA components
mkdir build && cd build
cmake .. && make -j$(nproc)

Quick Start with GUI

# Start the web-based GUI (fastest way to get started)
docking-at-home gui

# Or with Python
python -m docking_at_home.gui

# Open browser to http://localhost:8080

Quick Start Example (CLI)

from docking_at_home import DockingClient

# Initialize client (localhost mode)
client = DockingClient(mode="localhost")

# Submit docking job
job = client.submit_job(
    ligand="path/to/ligand.pdbqt",
    receptor="path/to/receptor.pdbqt",
    num_runs=100
)

# Monitor progress
status = client.get_status(job.id)

# Retrieve results
results = client.get_results(job.id)
print(f"Best binding energy: {results.best_energy} kcal/mol")

Running on Localhost

# Start server
docking-at-home server --port 8080

# In another terminal, run worker
docking-at-home worker --local

Citation

@software{docking_at_home_2025,
  title={Docking@HOME: A Distributed Platform for Molecular Docking},
  author={OpenPeer AI and Riemann Computing Inc. and Bleunomics and Andrew Magdy Kamal},
  year={2025},
  url={https://huggingface.co/OpenPeerAI/DockingAtHOME},
  license={GPL-3.0}
}

Component Citations

Please also cite the underlying technologies:

@article{morris2009autodock4,
  title={AutoDock4 and AutoDockTools4: Automated docking with selective receptor flexibility},
  author={Morris, Garrett M and Huey, Ruth and Lindstrom, William and Sanner, Michel F and Belew, Richard K and Goodsell, David S and Olson, Arthur J},
  journal={Journal of computational chemistry},
  volume={30},
  number={16},
  pages={2785--2791},
  year={2009}
}

@article{anderson2004boinc,
  title={BOINC: A system for public-resource computing and storage},
  author={Anderson, David P},
  journal={Grid Computing, 2004. Proceedings. Fifth IEEE/ACM International Workshop on},
  pages={4--10},
  year={2004},
  organization={IEEE}
}

Community & Support

• HuggingFace: huggingface.co/OpenPeerAI/DockingAtHOME
• Issues & Discussions: HuggingFace Discussions
• Email: [email protected]

Contributing

We welcome contributions from the community! Please see CONTRIBUTING.md

Areas for Contribution

• Algorithm improvements
• GPU optimization
• Web interface development
• Documentation
• Testing
• Bug reports
• Use case examples

License

This project is licensed under the GNU General Public License v3.0 - see LICENSE for details.

Individual components retain their original licenses:

• AutoDock: GNU GPL v2
• BOINC: GNU LGPL v3
• CUDPP: BSD License
• Decentralized Internet SDK: Various open-source licenses

Acknowledgments

• The AutoDock development team at The Scripps Research Institute
• UC Berkeley's BOINC project
• CUDPP developers and NVIDIA
• Lonero Team for the Decentralized Internet SDK
• OpenPeer AI for Cloud Agents framework
• All volunteer computing contributors worldwide

Version History

v1.0.0 (2025)

• Initial release
• AutoDock 4.2.6 integration
• BOINC distributed computing support
• CUDA/CUDPP GPU acceleration
• Decentralized Internet SDK integration
• Cloud Agents AI orchestration
• HuggingFace model card and datasets

---

Built with ❤️ by the open-source computational chemistry community

Repository: https://huggingface.co/OpenPeerAI/DockingAtHOME
Support: [email protected]