NVIDIA Hackathon

GPU-accelerated implementation of Scanpy's rank_genes_groups Wilcoxon rank-sum test. Developed for the NVIDIA x scverse Hackathon and for integration with rapids_singlecell ecosystem. The repo holds notebooks, scripts, and source modules for benchmarking and prototyping analyses.

Results

• 5-15x speedup of wilcoxon rank-sum test
• Perfect parity with Scanpy's rank_genes_groups(method = "wilcoxon") output
• Up to 2x faster than rapids_singlecell.tl.rank_genes_groups_logreg() for large (500k+ cell) datasets
• Automatic memory-based chunking strategy enables practically any GPU to be used. The GPU-accelerated Wilcoxon rank-sum was tested and validated on an RTX2080Ti and A100.

Dependencies

• All notebooks were run with NVIDIA's preconfigued NVIDIA Brev resources (https://github.com/NVIDIA-AI-Blueprints/single-cell-analysis-blueprint)
• Install the required packages from requirements.txt to run AstroWRLD tests and analysis.

Installation

uv pip install -r requirements.txt