netZooPy GPU Tutorial

A simple tutorial for running netZooPy (PANDA / LIONESS) with CuPy GPU acceleration on an HPC cluster (Open OnDemand, SLURM). The demo uses motif and PPI priors downloaded from GRAND; we also document sponge so you can use motif and PPI priors from sponge instead. Notebooks are compute-only; installs happen outside Jupyter.

This tutorial is an update of the netZooPy gpuPanda tutorial (Daniel Morgan, Channing Division of Network Medicine, Brigham and Women's Hospital / Harvard Medical School).

Table of Contents

• Prerequisites
• Install & Setup
• Usage
• Directory Structure
• Contributing
• License
• Authors
• References

Prerequisites

• Open OnDemand Jupyter app
• SLURM GPU partition
• NVIDIA GPU nodes (CUDA 12–compatible, e.g. A100)
• NVIDIA driver ≥ CUDA 12
• CUDA runtime provided via modules (no CUDA in conda; will set it up for this later)
• User Miniforge install

Assumed:

• cupy-cuda12x
• CUDA module close to CUDA/12.3.0

Install & Setup

Conda environment (for OOD Jupyter)

Use config/netzoo-gpu.yml or the block below:

name: netzoo-gpu
channels:
  - conda-forge

dependencies:
  - python=3.10
  - pip
  - numpy<1.24
  - pandas
  - psutil
  - s3fs
  - ipykernel

  - pip:
      - cupy-cuda12x

Notes: CuPy via pip to match system CUDA; CUDA from modules, not conda. NumPy pinned for netZooPy.

Set up the environment

On a hpc, load Miniforge and create/activate the environment, then register the Jupyter kernel.

module load miniforge/conda

conda env create -f config/netzoo-gpu.yml

conda activate netzoo-gpu

python -m ipykernel install --user \
  --name netzoo-gpu \
  --display-name "Python (netzoo-gpu)"

Install netZooPy from source

Install netZooPy into the netzoo-gpu environment from a user-writable source directory.

mkdir -p $HOME/src
cd $HOME/src

git clone --branch devel https://github.com/netZoo/netZooPy.git
cd netZooPy
pip install -e .

Alternative: sponge (motif and PPI priors)

You can use sponge to obtain motif and PPI priors instead of the GRAND downloads in the demo. For setup, see using sponge and config/sponge.yml.

Usage

1. Start Jupyter on a GPU node (Open OnDemand)
   Load CUDA before starting the Jupyter server.

   module load CUDA/12.3.0

   Select kernel: Python (netzoo-gpu).

2. Run the demo
   Open panda-gpu-demo.ipynb. It downloads LCL motif and PPI data from GRAND into data/, runs PANDA with GPU and precision options, and writes results to output/. To use motif and PPI priors from sponge instead, see Using sponge.

Troubleshooting

Verify Python path

import sys
sys.executable

Expected:

.../miniforge3/envs/netzoo-gpu/bin/python

Verify GPU access

import cupy as cp
cp.cuda.runtime.getDeviceCount()

If this fails:

• Confirm a GPU was requested in Open OnDemand
• Confirm CUDA/12.3.0 was loaded before Jupyter launch
• Restart the Jupyter server after module changes

Directory Structure

panda-gpu-tutorial/
├── config/           # Environment configs (conda, sponge)
│   ├── netzoo-gpu.yml
│   └── sponge.yml
├── data/             # Input data (e.g. from GRAND; see data/README.md)
├── docs/             # Project documentation (e.g. using-sponge.md)
├── output/           # PANDA and other pipeline outputs
├── src/              # Scripts (e.g. sponge.sh)
├── tests/            # Tests
├── panda-gpu-demo.ipynb
├── README.md
├── requirements.lock.txt
├── CONTRIBUTING.md
├── LICENSE
├── .github/
│   ├── ISSUE_TEMPLATE/
│   ├── PULL_REQUEST_TEMPLATE.md
│   └── workflows/
└── .gitignore

Contributing

We welcome contributions. See CONTRIBUTING.md for guidelines.

License

View the LICENSE for this project.

Authors

• Shaurita Hutchins email | PhD Candidate| sdhutchins

References

• netZooPy (PANDA, GPU): The Network Zoo: a multilingual package for the inference and analysis of gene regulatory networks. NAR Genomics and Bioinformatics 4(1), lqac002 (2022).
• sponge: sponge: reproducible Python environments for HPC. Bioinformatics 41(7), btaf320 (2025).