Scanpy quickstart
Run a GPU-accelerated single-cell analysis on the platform in a few steps.
You have single-cell RNA-seq counts in an AnnData file and you want clustered cells, a UMAP, and per-cluster marker genes. Scanpy on the platform runs the standard pipeline — QC → normalize → highly variable genes → PCA → neighbors → UMAP → Leiden → differential expression — on a GPU via rapids-singlecell, so even large datasets finish in minutes. This walk-through takes you from an .h5ad file to exported results.
Before you start
The platform's Scanpy job reads an AnnData .h5ad file from storage — either an S3 URI or an HTTPS URL ending in .h5ad. The full GPU pipeline runs on a T4 GPU; the batch-correction, annotation, and scoring extensions run on CPU. Runs are asynchronous with a 10-minute timeout.
Prepare your data
Upload an AnnData (.h5ad) file, or reuse one produced by an upstream job. If you have 10X Cell Ranger output or a count matrix, convert it to .h5ad first — that's the format the pipeline expects.
Choose a workflow
Pick what this run should do:
full(default) — the complete pipeline through clustering and differential expression.qc— quality control and filtering only.clustering— QC through Leiden clustering and UMAP, without marker genes.de— differential expression / marker genes.
Set quality-control thresholds
The defaults are a good starting point:
- Minimum genes per cell (
min_genes, default 200) — removes empty droplets. - Maximum mitochondrial percent (
max_mito_percent, default 5) — removes dying cells. - Highly variable genes (
n_top_genes, default 2000) — focuses downstream steps on informative genes. - Clustering resolution (
resolution, default 1.0) — lower values give fewer, larger clusters.
Run the analysis
Submit the job. The pipeline runs on GPU and reports progress through its phases. You can keep working; you'll be notified when it completes.
Review and export
When the job finishes, a result card shows the UMAP with clusters plus cell, gene, and cluster counts. Download the processed AnnData (with all embeddings and cluster labels), the marker-gene tables, and the publication-ready figures.
Start with the full workflow to see the whole picture, then re-run de or clustering at a different resolution if the cluster granularity doesn't match your biology — over-clustering is easier to merge than under-clustering is to split.
Read next
Was this page helpful?