| name | scrna-seq-pipeline |
| description | Scanpy-based single-cell RNA-seq analysis pipeline covering loading (10X, h5ad), QC, normalization, HVG selection, PCA/UMAP, Leiden clustering, differential expression, and batch correction (Harmony, scVI). Use when the user mentions single-cell, scRNA-seq, Scanpy, AnnData, UMAP, clustering, 10X, h5ad, leiden, highly variable genes, Harmony, scVI, cluster cells, find marker genes, filter low-quality cells, gene expression matrix, doublet removal, normalize counts, dimensionality reduction, cell clustering, QC single-cell, filtered_feature_bc_matrix, cellranger output, 10X h5, scrublet, cell ranger, count matrix, find cell types, batch effect, integration, neighborhood graph, or differential expression genes. |
| usage | Invoke when processing scRNA-seq data with Scanpy, including QC, normalization, clustering, DE, or batch correction. |
| version | 1.0.0 |
| validated_against | {"date":"2025-01-15T00:00:00.000Z","packages":{"scanpy":"1.10","anndata":"0.10","scvi-tools":"1.1"}} |
| tags | ["skill","category:pipeline","single-cell","scanpy","scrna-seq","hcls"] |
scRNA-seq Pipeline (Scanpy)
Overview
This skill produces a reproducible scRNA-seq analysis in Python using Scanpy and AnnData. It covers the standard workflow: load counts → QC filter → normalize/log → HVG → scale/PCA → neighbors/UMAP → Leiden → differential expression, plus batch correction with Harmony or scVI. Default parameters follow the Scanpy PBMC3k tutorial conventions and are safe starting points for most 10X Genomics datasets.
Output is a single .h5ad file with all intermediate embeddings, cluster labels, and DE results stored on the AnnData object.
Usage
Install dependencies:
pip install "scanpy>=1.10" anndata leidenalg python-igraph harmonypy scvi-tools
Run the pipeline end-to-end:
import scanpy as sc
import scanpy.external as sce
sc.settings.verbosity = 2
sc.settings.set_figure_params(dpi=100, facecolor="white")
adata = sc.read_10x_h5("filtered_feature_bc_matrix.h5")
adata.var_names_make_unique()
adata.var["mt"] = adata.var_names.str.startswith("MT-")
sc.pp.calculate_qc_metrics(adata, qc_vars=["mt"], percent_top=None, log1p=False, inplace=True)
sc.pp.filter_cells(adata, min_genes=200)
sc.pp.filter_genes(adata, min_cells=3)
adata = adata[adata.obs.n_genes_by_counts < 5000, :]
adata = adata[adata.obs.pct_counts_mt < 20, :].copy()
sc.pp.normalize_total(adata, target_sum=1e4)
sc.pp.log1p(adata)
adata.raw = adata
sc.pp.highly_variable_genes(adata, n_top_genes=2000, flavor="seurat")
adata = adata[:, adata.var.highly_variable].copy()
sc.pp.scale(adata, max_value=10)
sc.tl.pca(adata, n_comps=50, svd_solver="arpack")
sc.pp.neighbors(adata, n_neighbors=15, n_pcs=30)
sc.tl.umap(adata)
sc.tl.leiden(adata, resolution=0.5)
sc.tl.rank_genes_groups(adata, groupby="leiden", method="wilcoxon")
adata.write("result.h5ad", compression="gzip")
Batch correction — Harmony
sce.pp.harmony_integrate(adata, key="batch")
sc.pp.neighbors(adata, n_neighbors=15, n_pcs=30, use_rep="X_pca_harmony")
sc.tl.umap(adata)
sc.tl.leiden(adata, resolution=0.5)
Batch correction — scVI
scVI requires raw counts. Run it before normalization, or keep a raw-count layer.
import scvi
scvi.model.SCVI.setup_anndata(adata_raw, batch_key="batch")
model = scvi.model.SCVI(adata_raw, n_latent=30)
model.train(max_epochs=400, early_stopping=True)
adata.obsm["X_scVI"] = model.get_latent_representation()
sc.pp.neighbors(adata, n_neighbors=15, use_rep="X_scVI")
sc.tl.umap(adata)
sc.tl.leiden(adata, resolution=0.5)
Response Format
- Lead with the command or code the user needs — explain after
- Structure as: confirm inputs → working code → key parameters explained → gotchas
- One complete working example per task; do not show every alternative
- Keep code comments minimal and functional (what, not why-it-exists)
- Target: 50-100 lines of code with brief surrounding explanation
Core Concepts
- AnnData layout.
adata.X is the current expression matrix (cells × genes). adata.obs holds per-cell metadata (QC metrics, cluster labels, batch). adata.var holds per-gene metadata (HVG flags, mt). Embeddings live in adata.obsm (X_pca, X_umap, X_scVI). Neighbor graph and DE results live in adata.uns.
.raw snapshot. adata.raw = adata stores the full log-normalized matrix before HVG subsetting. Downstream plotting functions (sc.pl.umap(..., color="GENE"), sc.pl.rank_genes_groups) read from .raw so you can visualize any gene even after restricting adata to HVGs.
- QC thresholds are dataset-dependent.
n_genes ∈ [200, 5000] and pct_mt < 20 are reasonable defaults for human 10X PBMC-like data. Inspect distributions (sc.pl.violin) before committing. Use MT- for human and mt- for mouse.
- HVG flavor matters. Use
flavor="seurat" on log-normalized data (default). Use flavor="seurat_v3" on raw counts — it expects integer counts and will error or mislead on log data.
- Neighbors parameters.
n_neighbors controls local/global structure (10–50; larger = smoother UMAP). n_pcs should capture the elbow in sc.pl.pca_variance_ratio (typically 20–50). Leiden resolution controls cluster granularity (0.2–1.5; higher = more clusters).
- Batch correction choice. Harmony is fast, operates on PCA, and works well for mild-to-moderate batch effects. scVI is a deep generative model, needs raw counts and GPU for speed, and handles stronger technical variation and multi-modal designs.
Quick Reference
Loading
| Input | Call |
|---|
| 10X HDF5 | sc.read_10x_h5("filtered_feature_bc_matrix.h5") |
| 10X MTX dir | sc.read_10x_mtx("path/", var_names="gene_symbols", cache=True) |
| Existing AnnData | sc.read_h5ad("input.h5ad") |
| Always after load | adata.var_names_make_unique() |
Key parameter ranges
| Parameter | Default | Typical range | Notes |
|---|
filter_cells(min_genes=) | 200 | 100–500 | Drop empty/low-quality droplets |
filter_genes(min_cells=) | 3 | 3–10 | Drop rarely-expressed genes |
n_genes_by_counts < | 5000 | 2500–8000 | Upper bound flags doublets |
pct_counts_mt < | 20 | 5–25 | Human tissue dependent |
normalize_total(target_sum=) | 1e4 | 1e4 | CP10k; standard |
n_top_genes (HVG) | 2000 | 1000–5000 | More HVGs → more signal + noise |
scale(max_value=) | 10 | 10 | Clip extreme z-scores |
pca(n_comps=) | 50 | 30–100 | Pick elbow for downstream n_pcs |
neighbors(n_neighbors=) | 15 | 10–50 | Larger → smoother manifold |
neighbors(n_pcs=) | 30 | 20–50 | ≤ n_comps |
leiden(resolution=) | 0.5 | 0.2–1.5 | Higher → more clusters |
DE & inspection
sc.tl.rank_genes_groups(adata, groupby="leiden", method="wilcoxon")
sc.pl.rank_genes_groups(adata, n_genes=25, sharey=False)
import pandas as pd
result = adata.uns["rank_genes_groups"]
groups = result["names"].dtype.names
markers = pd.DataFrame({g: result["names"][g][:20] for g in groups})
Saving / loading
adata.write("result.h5ad", compression="gzip")
adata = sc.read_h5ad("result.h5ad")
Common Mistakes
-
Wrong: Subsetting to HVGs before saving adata.raw = adata
Right: Always set .raw on the log-normalized full matrix before adata = adata[:, adata.var.highly_variable].copy()
Why: Without .raw, you lose the ability to plot or score non-HVG genes in downstream analysis
-
Wrong: Using flavor="seurat_v3" on log-normalized data
Right: Match HVG flavor to data state: seurat_v3 requires raw integer counts; seurat (default) requires log-normalized data
Why: Wrong flavor produces nonsense HVG rankings or errors, corrupting all downstream steps
-
Wrong: Skipping adata.var_names_make_unique() after loading 10X data
Right: Call adata.var_names_make_unique() immediately after loading
Why: Duplicated gene symbols cause silent misbehavior in indexing, HVG selection, and DE analysis
-
Wrong: Running scVI on normalized/log-transformed data
Right: Feed scVI the raw-count AnnData and set batch_key via setup_anndata
Why: scVI models counts directly; normalized input violates its distributional assumptions and produces incorrect latent spaces
-
Wrong: Using MT- prefix for mouse mitochondrial genes (or mt- for human)
Right: Use MT- for human genes and mt- for mouse genes
Why: Wrong prefix makes pct_counts_mt always zero, rendering the QC filter a no-op
-
Wrong: Boolean-slicing AnnData without calling .copy()
Right: Always use .copy() after filtering: adata = adata[mask, :].copy()
Why: Without .copy(), the result is a view; subsequent in-place operations will warn or fail
References