RFdiffusion Open OnDemand Application

Description

Web interface for protein design using RFdiffusion on HPC clusters. Enables:

• De novo protein generation
• Motif scaffolding
• Binder design
• Symmetric oligomer generation
• Partial diffusion workflows

Features

Core Capabilities

• Five Design Modes:
  • Binder Design: Create protein binders with specified interfaces
  • Motif Scaffolding: Scaffold functional motifs into stable structures
  • Partial Diffusion: Controlled generation with adjustable noise steps
  • Unconditional Generation: Fully automated protein design
  • Symmetric Design: Generate symmetric protein assemblies

Advanced Functionality

• Hotspot Residue Specification: Paint key interaction residues
• Fold Conditioning: Guide generation with structural constraints
• Auxiliary Potentials: Incorporate energy-based guidance
• Inpainting: Combine sequence/structure constraints
• Flexible Peptide Handling: Design with dynamic binding partners

System Requirements

Hardware

• NVIDIA GPU (A100/V100 recommended)
• 60GB+ GPU memory
• 100GB+ temporary storage

Software

• Singularity 3.7+
• CUDA 11.6
• Python 3.9
• SE(3)-Transformer

Installation

1. Clone Repository:

git clone https://github.com/EpiGenomicsCode/RFDiffusion-OOD.git
cd RFDiffusion-OOD

2. Retrieve Singularity Container From ICDS:

rsync -avP path/torfdiffusion_container.sif .

Usage

Web Interface

1. Access Open OnDemand portal

2. Navigate to "RFdiffusion Protein Design"

3. Configure parameters:

   • Design Mode: Select workflow type
   • Input Structure: Upload PDB (if required)
   • Design Parameters:
     • Number of designs (1-100)
     • Diffusion timesteps (20-200)
     • Symmetry type (if applicable)
     • Potential guidance (optional)

4. Submit job through interactive form

Command Line Options

Example scaffolded binder design:

./scripts/run_inference.py \
  inference.input_pdb=input.pdb \
  inference.output_prefix=outputs/binder_design \
  scaffoldguided.scaffoldguided=True \
  'ppi.hotspot_res=[A59,A83,A91]' \
  inference.num_designs=10 \
  denoiser.noise_scale_ca=0

Common parameters:

• contigmap.contigs: Structural constraints
• potentials.guiding_potentials: Energy guidance
• symmetry.symmetry_type: Assembly symmetry
• partial_diffusion.partial_T: Noise steps

Output Structure

working_dir/
├── inputs/              # Uploaded PDB files
├── outputs/
│   ├── designs/         # Generated PDB structures
│   ├── scores/          # Design metrics
│   └── visualizations/  # 3D previews
├── logs/
│   ├── diffusion.log    # Full process log
│   └── status.json      # Progress tracking
└── schedules/           # Diffusion cache

Monitoring

Real-time tracking includes:

• Diffusion progress
• Energy landscape exploration
• Structural validation metrics
• Resource utilization

Troubleshooting

Issue	Solution
CUDA OOM	Reduce num_designs or use simpler contigs
Invalid PDB	Verify input structure with pdb-tools
Symmetry failures	Check symmetry parameters match input
Potential guidance conflicts	Adjust potential weights

License

MIT License

Acknowledgements

• RFdiffusion by Rosetta Commons
• NVIDIA SE3-Transformer
• This project is generously funded by Cornell University BRC Epigenomics Core Facility (RRID:SCR_021287), Penn State Institute for Computational and Data Sciences (RRID:SCR_025154) and Penn State University Center for Applications of Artificial Intelligence and Machine Learning to Industry Core Facility (RRID:SCR_022867)

Contact

• Technical Support: [email protected]
• Application Maintainer: [email protected]