What problem does it solve? CPU-based Scanpy pipelines become prohibitively slow on large single-cell and spatial datasets exceeding 500K cells, and this Skill provides a complete GPU-accelerated workflow on FGCZ infrastructure using rapids_singlecell and NVIDIA L40S GPUs. ## Core Features & Use Cases - Reproducible GPU Environment: Pixi-based environment definition with the full RAPIDS stack (cupy, cuml, cugraph, cuvs) plus rapids-singlecell, scanpy, and Marimo notebooks. - Seurat to h5ad Conversion: R script that converts large Seurat v5 multi-layer objects to h5ad format, preserving metadata and embeddings for AnnData-based GPU analysis. - GPU Preprocessing and Clustering: RMM managed-memory setup, normalization, HVG selection, PCA, neighbor search (brute or ivfflat), UMAP, and Leiden clustering entirely on GPU, with SBATCH scripts for SLURM GPU partition submission. - Use Case: A researcher with a 2.4M-cell merged Xenium dataset submits a GPU job that runs the full preprocessing-to-clustering pipeline in about 4 minutes and exports an interactive Marimo HTML report. ## Quick Start Ask the agent to set up a rapids_singlecell GPU analysis environment with Pixi and convert your Seurat object to h5ad for GPU-accelerated clustering on the FGCZ cluster.
