Protein Design Suite

Overview

An Open OnDemand Batch Connect app that provides a web-based interface for running de novo protein design jobs on GPU nodes. The app supports two complementary tools and handles all backend configuration, job submission, and real-time progress monitoring automatically.

• RFdiffusion — diffusion-based de novo protein design (binder design, motif scaffolding, partial diffusion, unconditional generation, symmetric oligomers)
• BoltzGen — universal binder design for proteins, peptides, nanobodies, antibodies, and small molecules

Batch Connect template: basic | Scheduler: Slurm | Container runtime: Singularity / Apptainer

---

How It Looks

BoltzGen input form

BoltzGen — part 1	BoltzGen — part 2

RFdiffusion input form

RFdiffusion — part 1	RFdiffusion — part 2

Progress during and after a run

BoltzGen running (live log + step tracker)	RFdiffusion completed

---

Features

• Unified form — select RFdiffusion or BoltzGen from a single dropdown; irrelevant fields hide automatically
• RFdiffusion modes: binder design (PPI), motif scaffolding, partial diffusion, unconditional generation, symmetric oligomers
• BoltzGen protocols: protein-anything, peptide-anything, nanobody-anything, antibody-anything, protein→small molecule, protein redesign
• Real-time progress bar — startup milestones, per-design counters (RFdiffusion) and per-pipeline-step tracking (BoltzGen)
• Live log panel — streams the job log directly in the browser
• Automatic GPU resource management — GPU constraint enforced to avoid incompatible hardware
• Results as PDB / CIF files with full trajectory information

---

Prerequisites

Open OnDemand

• OOD v3 or v4 (tested on OOD v4)
• Slurm scheduler

GPU Requirements

The containers require PyTorch ≥ sm_70. On ROAR Collab the SLURM templates already set #SBATCH --constraint=v100|a100 — this has been tested and works. If your cluster uses different GPU labels, update that line in template/boltzgen.sh and template/rfdiffusion.sh. If you need to run on older GPUs, try rebuilding the container from the ProteinDesign-Containers definition files with a compatible PyTorch version.

Singularity Containers and Models

Build or pull the containers using the definition files in ProteinDesign-Containers:

Image	File
rfdiffusion_x86.sif	rfdiffusion/rfdiffusion_x86.def
boltzgen_x86.sif	boltzgen/boltzgen_x86.def

BoltzGen downloads model weights automatically from Hugging Face on first run and caches them in the directory bound to /models inside the container. Set BOLTZ_MODEL_DIR in template/rfdiffusion_env.sh to a persistent path with sufficient storage (~6 GB).

---

Installation

1. Clone this repository into your OOD dev apps directory:

   cd ~/ondemand/dev
   git clone https://github.com/EpiGenomicsCode/ProteinDesign-Suite RFDiff

2. Build or pull the Singularity containers (see ProteinDesign-Containers).

3. Edit template/rfdiffusion_env.sh and set the paths for your site:

   CONTAINER_BASE="/path/to/your/containers"   # directory with .sif files
   BOLTZ_MODEL_DIR="${CONTAINER_BASE}/boltzgen_models"

4. Verify the app loads in your OOD dashboard under Interactive Apps → Protein Engineering.

---

Configuration

template/rfdiffusion_env.sh

All site-specific paths live here — this is the only file that needs editing for a new deployment.

Variable	Description	Default
CONTAINER_BASE	Directory containing .sif files	(set by admin)
BOLTZ_MODEL_DIR	BoltzGen model cache directory	${CONTAINER_BASE}/boltzgen_models
DIFFUSION_SIF	RFdiffusion container path	${CONTAINER_BASE}/rfdiffusion_x86.sif
BOLTZ_SIF	BoltzGen container path	${CONTAINER_BASE}/boltzgen_x86.sif

form.yml.erb — key attributes

Attribute	Widget	Description	Default
app_type	select	RFdiffusion or BoltzGen	rfdiffusion
auto_accounts	select	GPU-capable Slurm account	(dynamic)
working_directory	path_selector	Output directory	/scratch/<user>
design_mode	select	RFdiffusion mode	binder
num_designs	number	Designs to generate	10
boltz_protocol	select	BoltzGen protocol	protein-anything
boltz_num_designs	number	BoltzGen intermediate designs	100
boltz_budget	number	Final filtered design count	10

submit.yml.erb

The wrapper job runs on the basic partition with the open account. The actual compute job (RFdiffusion or BoltzGen) is submitted separately inside before.sh.erb with GPU resources on the standard partition. Update the partition and account names to match your cluster.

---

Usage

1. Open the OOD dashboard and navigate to Interactive Apps → Protein Engineering → Protein Design Suite
2. Select the tool (RFdiffusion or BoltzGen)
3. Enter your GPU account and working directory
4. Fill in the tool-specific parameters:
   • RFdiffusion: choose design mode, provide a target PDB (required for all modes except unconditional), set length / contigs
   • BoltzGen: choose protocol, provide a target structure file, set binder length range and budget
5. Submit — the app launches a wrapper job that stages inputs and submits the compute job
6. Monitor progress in the OOD session panel: live step tracker, percentage bar, and streaming log

Output structure

<working_directory>/
└── pp<session_id>/
    ├── input/            # Staged inputs and generated SLURM scripts
    ├── structure/        # Design outputs (PDB / CIF files)
    └── logs/             # Job logs (streamed live in the browser)

---

Troubleshooting

Symptom	Likely cause	Fix
"Permission denied" on mkdir	Working directory path not writable	Use /scratch/<user> or another writable path
"undefined method" on staging	Form field missing from form.yml.erb	Pull latest — all context fields are now declared
Progress bar stuck at 0% "Starting up"	Old staged script.sh in session	Cancel and resubmit; templates are updated
BoltzGen: CUDA error / no kernel image	Job landed on a P100 GPU	Ensure `--constraint=v100
RFdiffusion: FileNotFoundError for PDB	Container path mismatch	Check inference.input_pdb path in template/rfdiffusion.sh
BoltzGen step 4 AssertionError	Stale output dir with --reuse	Run without --reuse (default) for fresh jobs

---

Testing

Tested on ICDS ROAR Collab (Penn State) with:

• OOD v4
• Slurm
• Singularity / Apptainer
• V100 and A100 GPUs

All five RFdiffusion modes (binder, scaffold, partial, unconditional, symmetric) and the BoltzGen protein-anything protocol have been run end-to-end through the form.

---

Contributing

For bugs or feature requests, open an issue.

---

References

• RFdiffusion — Watson et al., Nature 2023
• BoltzGen — universal binder design
• ProteinDesign-Containers — Singularity definition files
• Open OnDemand — HPC portal framework
• OOD Batch Connect app development docs

---

License

MIT — see LICENSE

Acknowledgements

Supported by 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 (AIMI) (RRID:SCR_022867).

Contact

• Technical support: vinaysmathew@psu.edu
• ICDS support: icds@psu.edu