Orchestrates the end-to-end single-cell RNA-seq pipeline from 10x Cell Ranger output to annotated cell types, chaining ambient-RNA removal, doublet detection, MAD-adaptive QC, normalization, integration, clustering, marker annotation, and (separately) pseudobulk DE + differential abundance. Use when honoring the made-once counting commitments (reference build/Ensembl vintage, --include-introns, cell-calling, feature namespace), ordering correct-then-detect-then-normalize (ambient before doublet before normalize), running per-sample QC before merge, integrating-then-clustering (never testing on integrated values), aggregating to pseudobulk for condition DE instead of cells-as-replicates, or pairing DE with differential abundance. Hands mechanism to the single-cell component skills; not a re-teach of any single step.
Orchestrates the end-to-end single-cell RNA-seq pipeline from 10x Cell Ranger output to annotated cell types, chaining ambient-RNA removal, doublet detection, MAD-adaptive QC, normalization, integration, clustering, marker annotation, and (separately) pseudobulk DE + differential abundance. Use when honoring the made-once counting commitments (reference build/Ensembl vintage, --include-introns, cell-calling, feature namespace), ordering correct-then-detect-then-normalize (ambient before doublet before normalize), running per-sample QC before merge, integrating-then-clustering (never testing on integrated values), aggregating to pseudobulk for condition DE instead of cells-as-replicates, or pairing DE with differential abundance. Hands mechanism to the single-cell component skills; not a re-teach of any single step.
Before using code patterns, verify installed versions match. If versions differ:
Python: pip show <package> then help(module.function) to check signatures
R: packageVersion('<pkg>') then ?function_name to verify parameters
If code throws ImportError, AttributeError, or TypeError, introspect the installed
package and adapt the example to match the actual API rather than retrying.
Note: the Cell Ranger filtered_feature_bc_matrix is cell-CALLING only, NOT ambient-corrected; recovering low-RNA cells or running SoupX/CellBender needs the RAW matrix. --include-introns is default TRUE since Cell Ranger 7 (essential for nuclei). gex_only=True/default silently drops Antibody Capture/CRISPR/HTO features. Confirm in-tool before quoting.
Single-Cell RNA-seq Pipeline
"Analyze my single-cell RNA-seq data from counts to cell types" -> Orchestrate QC filtering, normalization (scanpy/Seurat), batch integration (scVI/Harmony), clustering, marker detection, cell type annotation, and trajectory inference.
This is a workflow skill: it owns the chaining decisions and hand-offs, not the internals of any one step. Every step below cross-references the component skill that teaches its mechanism.
Made-once commitments (decided at cellranger count, inherited downstream)
Every gene ID/marker lookup/cross-dataset merge; join on gene_ids (stable), not symbols (lossy across releases)
--include-introns (default TRUE since CR 7)
Total UMI/cell and snRNA-vs-scRNA comparability; nuclei need introns; mixing intron-on/off samples is a hidden batch axis
Cell-vs-empty-droplet calling
The filtered matrix is cell-CALLING only, NOT ambient-corrected; SoupX/CellBender and low-RNA-cell rescue need the RAW matrix (filtered-only is irreversible loss)
Feature/barcode namespace (gex_only)
gex_only=True silently drops ADT/HTO/CRISPR features; set False and split by for CITE-seq/hashed pools
feature_types
Pipeline orchestration: the ordering decisions that make or break the result
Stage order is not arbitrary; getting it wrong silently corrupts every downstream step. The cross-cutting decisions this pipeline must get right:
Correct-then-detect-then-normalize, never reorder. Ambient-RNA removal (SoupX/CellBender) reads the RAW droplet matrix and MUST precede doublet detection and the final QC thresholds (ambient inflation distorts both doublet scores and per-cell QC); doublet detection precedes integration (doublets seed fake intermediate clusters). See single-cell/preprocessing.
Multiplexed (hashed) pools are demultiplexed FIRST. When samples are pooled with cell hashing (HTO/MULTI-seq) or genotype multiplexing, assign each cell to its sample of origin and remove cross-sample doublets before any per-sample step; the per-sample QC and doublet operations below assume the pool is already split by sample. See single-cell/hashing-demultiplexing.
Per-sample QC and doublet detection come BEFORE any merge or integration. Ambient-RNA cleanup (SoupX/CellBender), mito/gene-count filtering, and doublet calling (expected rate ~0.8% per 1,000 recovered cells) are per-capture operations; running them on a pooled object lets one lane's artifacts contaminate the shared null and lets surviving doublets form fake intermediate clusters. See single-cell/preprocessing and single-cell/doublet-detection.
Integrate, THEN cluster, THEN annotate - never reorder. For multi-sample designs, batch-correct on a shared embedding (Harmony/scVI/RPCA) and cluster on the corrected graph; clustering before correction makes clusters track samples or lanes instead of biology, and annotation has nothing to label until clusters exist. See single-cell/batch-integration, single-cell/clustering, single-cell/cell-annotation.
Over-integration erases biology. Batch-mixing metrics (kBET, iLISI) are maximized by destroying structure, so a method that flattens real cell states scores perfectly; pair every batch metric with a bio-conservation metric and re-inspect rare populations after correction. Batch-corrected expression is for embedding and clustering only - never feed it to differential expression. See single-cell/batch-integration.
Condition-level DE must use pseudobulk, not cells-as-replicates. Testing thousands of cells per donor as independent replicates is pseudoreplication and inflates false positives by orders of magnitude (Squair 2021); aggregate RAW counts per sample x cell type and hand them to DESeq2/edgeR/limma-voom. Step 7 marker p-values are descriptive ranking for labeling, not condition inference. See single-cell/markers-annotation and differential-expression/deseq2-basics.
Composition shifts are tested separately and masquerade as DE. A cluster-level expression change between conditions can be a pure proportion shift (a mixed cluster's substates re-balance with no gene changing per cell), invisible if only DE is run; always pair condition DE with a differential-abundance test (Milo cluster-free, or scCODA/sccomp/propeller cluster-based). See single-cell/differential-abundance.
The Seurat and Scanpy paths below are written single-sample for clarity. A multiplexed design demultiplexes the pool first (single-cell/hashing-demultiplexing); a multi-sample design then inserts per-sample QC and doublet removal, a merge, and an integration step between normalization (Step 4) and dimensionality reduction (Step 5); the rest of the order is unchanged.
Workflow Overview
10X data (RAW + filtered_feature_bc_matrix)
|
v
[0. Ambient removal] ---> SoupX/CellBender on RAW (per sample) (single-cell/preprocessing)
|
v
[1. Load Data] ---------> Read10X / read_10x_h5
|
v
[2. QC + Doublets] -----> MAD-adaptive nFeature/percent.mt; scDblFinder/Scrublet (per sample, before merge)
|
v
[3. Normalization] -----> SCTransform or LogNormalize
|
v
[4. HVG Selection] -----> FindVariableFeatures
|
v
[5. Dim Reduction] -----> PCA -> UMAP
|
v
[6. Clustering] --------> FindNeighbors -> FindClusters
|
v
[7. Markers] -----------> FindAllMarkers
|
v
[8. Annotation] --------> Manual or automated
|
v
Annotated Seurat/AnnData object
Primary Path: Seurat (R)
The Seurat/Scanpy paths below start from the filtered matrix for clarity. In practice, run ambient-RNA removal FIRST on the RAW droplet matrix per sample (SoupX needs raw+filtered+clusters; CellBender folds calling+denoise and is best for snRNA/high-ambient) — pick ONE (double-correction over-strips). See single-cell/preprocessing.
Regressing out percent.mt (or nCount/cell-cycle) is NOT reflexive: those covariates are confounded with real cell state and regressing them can erase biology, so only pass vars.to.regress for a covariate verified not confounded with the signal of interest. See single-cell/preprocessing.
# Set identity to chosen resolution
Idents(seurat_obj)<-'SCT_snn_res.0.4'# Find markers for all clusters
markers <- FindAllMarkers(seurat_obj, only.pos =TRUE, min.pct =0.25, logfc.threshold =0.25)# Top markers per cluster
top_markers <- markers %>%
group_by(cluster)%>%
slice_max(n =10, order_by = avg_log2FC)# Visualize top markers
DoHeatmap(seurat_obj, features = top_markers$gene)+ NoLegend()
Step 8: Cell Type Annotation
# Manual annotation based on known markers# Example for PBMC data:
cluster_annotations <-c('0'='CD4 T cells','1'='CD14 Monocytes','2'='B cells','3'='CD8 T cells','4'='NK cells','5'='CD16 Monocytes','6'='Dendritic cells')
seurat_obj$cell_type <- cluster_annotations[as.character(Idents(seurat_obj))]# Final UMAP
DimPlot(seurat_obj, reduction ='umap', group.by ='cell_type', label =TRUE)# Save object
saveRDS(seurat_obj,'seurat_annotated.rds')
Alternative Path: Scanpy (Python)
import scanpy as sc
import numpy as np
# Load 10X data
adata = sc.read_10x_h5('filtered_feature_bc_matrix.h5')
adata.var_names_make_unique()
# QC metrics
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)
# Filter (flat cutoffs are illustrative; prefer MAD-adaptive, tissue-aware thresholds, see single-cell/preprocessing)
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, :]
# Doublet detection (rate from recovered cells, not the 0.05 placeholder, see single-cell/doublet-detection)
expected_rate = 0.008 * adata.n_obs / 1000
sc.pp.scrublet(adata, expected_doublet_rate=expected_rate)
adata = adata[~adata.obs['predicted_doublet'], :]
# Normalize and HVGs
sc.pp.normalize_total(adata, target_sum=1e4)
sc.pp.log1p(adata)
sc.pp.highly_variable_genes(adata, n_top_genes=2000)
# PCA, neighbors, UMAP -- stash log-normalized values in .raw BEFORE scaling, or rank_genes_groups# later computes logfoldchanges from z-scores (negative means -> NaN/garbage LFCs)
adata.raw = adata
adata = adata[:, adata.var.highly_variable]
sc.pp.scale(adata, max_value=10)
sc.tl.pca(adata, n_comps=50)
sc.pp.neighbors(adata, n_neighbors=15, n_pcs=30)
sc.tl.umap(adata)
# Clustering (pin the backend for reproducibility; default flips to igraph)
sc.tl.leiden(adata, resolution=0.5, flavor='igraph', n_iterations=2, directed=False)
# Markers
sc.tl.rank_genes_groups(adata, 'leiden', method='wilcoxon')
sc.pl.rank_genes_groups(adata, n_genes=10, sharey=False)
# Save
adata.write('scanpy_annotated.h5ad')
Parameter Recommendations
Step
Parameter
Recommendation
QC
min.features
200-500
QC
max.features
2500-5000 (depends on data)
QC
percent.mt
<10-20% (tissue-dependent)
SCTransform
vars.to.regress
none by default; only a validated non-confounded covariate
PCA
npcs
30-50
UMAP
dims
15-30 (check elbow plot)
Clustering
resolution
0.4-0.8 (start with 0.5)
QC cutoffs above are illustrative starting points, not universal thresholds: set min/max features and mito % per dataset from the data (MAD-adaptive, tissue-aware) rather than porting flat values, since healthy mito fraction varies by tissue. See single-cell/preprocessing.
Ambient markers everywhere (e.g. Hb across all PBMCs)
Ran SoupX/CellBender on the FILTERED matrix (soup already removed)
Run ambient removal on the RAW droplet matrix, before QC
Fake intermediate cell states
Doublets removed AFTER clustering/integration
Call doublets per sample before merge/normalize
Clusters track sample/lane, not biology
Clustered before integration
Integrate -> cluster -> annotate
Inflated DE, thousands of false positives
Tested cells as replicates (on integrated values)
Pseudobulk RAW counts per sample x cell-type -> DESeq2/edgeR/limma (Squair 2021)
"DE change" is really a proportion shift
Only ran DE, not abundance
Pair with differential abundance (Milo/scCODA/propeller)
Biology collapses to one blob
Reflexively regressed total_counts/cell-cycle
Regress only a validated, non-confounded covariate
CITE-seq/hashtag features vanish
gex_only=True default
Set False; split by feature_types
References
Heumos L, Schaar AC, Lance C, et al (2023) Best practices for single-cell analysis across modalities. Nature Reviews Genetics 24:550-572. DOI 10.1038/s41576-023-00586-w. (canonical pipeline order.)
Squair JW, Gautier M, Kathe C, et al (2021) Confronting false discoveries in single-cell differential expression. Nature Communications 12:5692. DOI 10.1038/s41467-021-25960-2. (cells-as-replicates is pseudoreplication; pseudobulk.)
Fleming SJ, Chaffin MD, Arduini A, et al (2023) Unsupervised removal of systematic background noise from droplet-based single-cell experiments using CellBender. Nature Methods 20:1323-1335. DOI 10.1038/s41592-023-01943-7. (ambient removal on the RAW matrix.)
Related Skills
database-access/geo-data - Resolve GSE to SRA; detect SuperSeries before processing
database-access/sra-data - Download 10x records with --include-technical for barcodes/UMIs
single-cell/data-io - Loading 10X, h5ad, RDS, and h5mu formats
