Orchestrates the end-to-end spatial transcriptomics pipeline from Space Ranger / vendor output to spatial domains and statistics, branching FIRST on platform class (imaging in-situ Xenium/MERFISH/CosMx vs sequencing/capture Visium/Visium HD/Slide-seq). Use when deciding segmentation-vs-deconvolution and the QC floors from the platform class, committing the coordinate/image-registration frame and panel identity, deconvolving multi-cell spots against an annotated scRNA reference (never relabeling spot clusters as cell types), building the spatial neighbor graph on PHYSICAL not expression space, gating spatially-variable genes on FDR, or using a real domain method (BANKSY/BayesSpace/STAGATE) rather than clustering the spatial graph alone. Hands off deconvolution and cell-cell communication to the component skills; not a re-teach of any single step.
Orchestrates the end-to-end spatial transcriptomics pipeline from Space Ranger / vendor output to spatial domains and statistics, branching FIRST on platform class (imaging in-situ Xenium/MERFISH/CosMx vs sequencing/capture Visium/Visium HD/Slide-seq). Use when deciding segmentation-vs-deconvolution and the QC floors from the platform class, committing the coordinate/image-registration frame and panel identity, deconvolving multi-cell spots against an annotated scRNA reference (never relabeling spot clusters as cell types), building the spatial neighbor graph on PHYSICAL not expression space, gating spatially-variable genes on FDR, or using a real domain method (BANKSY/BayesSpace/STAGATE) rather than clustering the spatial graph alone. Hands off deconvolution and cell-cell communication to the component skills; not a re-teach of any single step.
Reference examples tested with: Space Ranger 4.1+ (Visium HD; nucleus/cell segmentation in the count pipeline since v4.0), scanpy 1.10+, squidpy 1.3+, spatialdata-io current (imaging platforms), matplotlib 3.8+, numpy 1.26+
Before using code patterns, verify installed versions match. If versions differ:
Python: pip show <package> then help(module.function) to check signatures
If code throws ImportError, AttributeError, or TypeError, introspect the installed
package and adapt the example to match the actual API rather than retrying.
Note: squidpy.read provides visium/vizgen/nanostring only — there is NO sq.read.xenium; imaging platforms load via spatialdata_io (returns a SpatialData object preserving the molecule table). Visium HD default bin is 8 µm. Confirm in-tool before quoting.
Spatial Transcriptomics Pipeline
"Analyze my spatial transcriptomics data end-to-end" -> Orchestrate data loading (squidpy/scanpy), QC, normalization, spatial neighbor analysis, spatial statistics, spatial domain detection, and tissue visualization. Composition estimation (deconvolution, spatial-deconvolution) and cell-cell communication (spatial-communication) are deliberately separate steps -- this pipeline hands off to those skills rather than inlining them.
This is a workflow skill: it owns the chaining decisions and hand-offs, not the internals of any one step.
Made-once commitments
Commitment
Consequence inherited downstream
Platform class (imaging vs sequencing)
EVERY downstream choice: segment-vs-deconvolve, discovery-vs-classification, QC floors, panel-bounded-vs-whole-transcriptome
Coordinate system + image-registration frame
All spatial neighbors/overlays/niches; a wrong registration frame silently misplaces every spot relative to histology
Panel identity (targeted vs whole-transcriptome; FFPE probe vs FF poly-A)
What "gene absent" means: on a targeted panel absence = "not in panel", not "not expressed"; RIN (FF) vs DV200 (FFPE) QC metric switch
Whether to DECONVOLVE (spot >> cell), SEGMENT/bin-up (spot << cell), or neither
Segmentation policy (imaging: Baysor / Cellpose / vendor Xenium; Visium HD bin-to-cell: Space Ranger v4+)
Every cell x gene value; segmentation is the DOMINANT imaging error source and over-expansion manufactures cross-type DE
The platform-class fork (decide first)
This pipeline branches on platform class before any step. Sequencing/capture data (Visium, Visium HD, Slide-seq, Stereo-seq) are spot/bin MIXTURES of cells: QC on spot counts, normalize knowing that library size partly carries cellularity, then DECONVOLVE composition (spatial-deconvolution) rather than read a spot as one cell. Imaging/in-situ data (Xenium, MERFISH, CosMx) are single molecules: SEGMENT cells first (image-analysis), apply low-count-aware QC floors (an scRNA min_counts=500 deletes nearly every real imaging cell, whose vector is tens-to-low-hundreds of transcripts), drop on negative-control probe rate, and SKIP deconvolution. The Squidpy+Scanpy path below is written for Visium; the imaging branch is flagged at each step.
Workflow Overview
Spatial data (Space Ranger output)
|
v
[1. Load Data] ---------> Read Visium/Xenium
|
v
[2. QC & Preprocessing] -> Filter, normalize
|
v
[3. Clustering] --------> Standard scRNA-seq clustering
|
v
[4. Spatial Analysis] --> Neighbors, statistics
|
v
[5. Domain Detection] --> Spatial domains
|
v
[6. Visualization] -----> Spatial plots
|
v
Annotated spatial data
Primary Path: Squidpy + Scanpy
Step 1: Load Data
import scanpy as sc
import squidpy as sq
import numpy as np
import matplotlib.pyplot as plt
# Load Visium data (Space Ranger output). squidpy.read provides only visium,# vizgen, and nanostring -- there is NO sq.read.xenium.
adata = sq.read.visium('spaceranger_output/')
# For Xenium and other imaging platforms use spatialdata_io, which returns a# SpatialData object preserving the per-transcript molecule table (the cell# matrix is one table inside it). See spatial-data-io.# import spatialdata_io as sdio# sdata = sdio.xenium('xenium_output/')# adata = sdata.tables['table'] # segmentation-derived cell matrixprint(f'Loaded: {adata.n_obs} spots/cells, {adata.n_vars} genes')
Step 2: Quality Control
# QC metrics. Mito genes are present on Visium but usually OFF-PANEL for imaging# platforms, so guard the mito calculation rather than assuming MT- genes exist.
has_mito = adata.var_names.str.startswith('MT-').any()
if has_mito:
adata.var['mt'] = adata.var_names.str.startswith('MT-')
sc.pp.calculate_qc_metrics(adata, qc_vars=['mt'], inplace=True)
else:
sc.pp.calculate_qc_metrics(adata, inplace=True)
# Always inspect QC SPATIALLY (a gradient across the section is a technical# artifact, not biology); violins alone hide it.
sc.pl.spatial(adata, color='total_counts', show=False)
plt.savefig('qc_spatial.pdf')
# Filter. These floors are VISIUM defaults (spot = 1-10-cell mixture) and are# tissue-dependent. For IMAGING data use low-count-aware floors (~10 transcripts# per cell, NOT 500) or aggressive filtering deletes nearly every real cell and# preferentially removes small cells (lymphocytes), biasing composition.
sc.pp.filter_cells(adata, min_counts=500)
sc.pp.filter_genes(adata, min_cells=10)
if has_mito:
adata = adata[adata.obs.pct_counts_mt < 25, :]
print(f'After QC: {adata.n_obs} spots/cells')
Step 3: Normalization and Clustering
# Store raw counts
adata.layers['counts'] = adata.X.copy()
# Normalize. In spatial data library size partly CARRIES BIOLOGY (Visium total# counts confound with cells-per-spot and cellularity; imaging total counts with# cell size), so total-count normalization is a Visium starting point, not a# universal default -- for imaging consider cell volume/area normalization and# see spatial-preprocessing before dividing library size out.
sc.pp.normalize_total(adata, target_sum=1e4)
sc.pp.log1p(adata)
# HVGs
sc.pp.highly_variable_genes(adata, n_top_genes=2000)
# PCA and clustering
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)
sc.tl.leiden(adata, resolution=0.5, flavor='igraph', n_iterations=2, directed=False)
# Visualize clusters in space. On Visium these spot clusters are REGIONS/niches,# NOT cell types -- a spot is a 1-10-cell mixture, so recovering cell-type# composition needs deconvolution (spatial-deconvolution), not clustering.
sc.pl.spatial(adata, color='leiden', spot_size=1.5)
plt.savefig('clusters_spatial.pdf')
Step 4: Spatial Analysis
# Build spatial neighbors graph. Visium is a hex lattice -> coord_type='grid'# (n_neighs=6); 'generic' kNN is for imaging point clouds. See spatial-neighbors.
sq.gr.spatial_neighbors(adata, coord_type='grid', n_neighs=6)
# Neighborhood enrichment. The Squidpy permutation null only tests "more adjacent# than complete spatial randomness" -- two abundant types sharing a compartment# pass trivially. A positive z is NOT a specific A-B interaction; demand a# conditional/toroidal null before claiming affinity. See spatial-statistics.
sq.gr.nhood_enrichment(adata, cluster_key='leiden')
sq.pl.nhood_enrichment(adata, cluster_key='leiden')
plt.savefig('nhood_enrichment.pdf')
# Co-occurrence analysis
sq.gr.co_occurrence(adata, cluster_key='leiden')
sq.pl.co_occurrence(adata, cluster_key='leiden')
plt.savefig('co_occurrence.pdf')
# Spatially variable genes. Gate on FDR, not raw I; and a top-Moran gene is# usually a marker of a spatially-clustered cell TYPE (composition), not a gene# regulated WITHIN a type -- intersect with non-HVG to find the latter. See# spatial-statistics.
sq.gr.spatial_autocorr(adata, mode='moran', n_perms=100, n_jobs=4)
moran = adata.uns['moranI']
svg = moran[moran['pval_norm_fdr_bh'] < 0.05].sort_values('I', ascending=False)
print('Spatially autocorrelated genes (FDR<0.05):', svg.head(10).index.tolist())
Step 5: Domain Detection
# Spatial domain detection. Clustering the spatial graph topology ALONE (below)# is a quick proxy, NOT a real domain method -- it ignores expression and carries# none of the over-smoothing / spatial-weight-knob / k-as-biological-choice# framing. For real domains use BANKSY (lambda ~0.8), BayesSpace, or STAGATE and# tune the spatial weight. See spatial-domains.
sq.gr.spatial_neighbors(adata, coord_type='grid', n_neighs=6)
sc.tl.leiden(adata, resolution=0.3, key_added='spatial_domains',
adjacency=adata.obsp['spatial_connectivities'],
flavor='igraph', n_iterations=2, directed=False)
# Visualize domains
sc.pl.spatial(adata, color='spatial_domains', spot_size=1.5)
plt.savefig('spatial_domains.pdf')
# Compare transcriptomic vs spatial clusters
sc.pl.spatial(adata, color=['leiden', 'spatial_domains'], ncols=2)
plt.savefig('clusters_comparison.pdf')
Step 6: Visualization
# Gene expression in space
genes = ['EPCAM', 'VIM', 'PTPRC', 'COL1A1']
sc.pl.spatial(adata, color=genes, ncols=2, spot_size=1.5, cmap='viridis')
plt.savefig('marker_genes_spatial.pdf')
# Cluster markers in space. On Visium these are markers of spot REGIONS (mixtures),# not of pure cell types; for cell-type-level signal deconvolve first.
sc.tl.rank_genes_groups(adata, 'leiden', method='wilcoxon')
sc.pl.rank_genes_groups_dotplot(adata, n_genes=5)
plt.savefig('cluster_markers.pdf')
# Save
adata.write('spatial_analyzed.h5ad')
Complete Workflow Script
import scanpy as sc
import squidpy as sq
import matplotlib.pyplot as plt
import os
# Configuration
data_dir = 'spaceranger_output'
output_dir = 'spatial_results'
os.makedirs(output_dir, exist_ok=True)
os.makedirs(f'{output_dir}/plots', exist_ok=True)
# Loadprint('Loading data...')
adata = sq.read.visium(data_dir)
print(f'Loaded: {adata.n_obs} spots, {adata.n_vars} genes')
# QC (Visium defaults; for imaging use low-count-aware floors and skip mito)print('QC filtering...')
has_mito = adata.var_names.str.startswith('MT-').any()
if has_mito:
adata.var['mt'] = adata.var_names.str.startswith('MT-')
sc.pp.calculate_qc_metrics(adata, qc_vars=['mt'], inplace=True)
else:
sc.pp.calculate_qc_metrics(adata, inplace=True)
sc.pp.filter_cells(adata, min_counts=500)
sc.pp.filter_genes(adata, min_cells=10)
if has_mito:
adata = adata[adata.obs.pct_counts_mt < 25, :]
print(f'After QC: {adata.n_obs} spots')
# Normalize and clusterprint('Processing...')
adata.layers['counts'] = adata.X.copy()
sc.pp.normalize_total(adata, target_sum=1e4)
sc.pp.log1p(adata)
sc.pp.highly_variable_genes(adata, n_top_genes=2000)
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.leiden(adata, resolution=0.5, flavor='igraph', n_iterations=2, directed=False)
# Spatial analysis (Visium hex -> coord_type='grid'; nhood z and top-Moran genes# need the caveats from Step 4 before interpretation)print('Spatial analysis...')
sq.gr.spatial_neighbors(adata, coord_type='grid', n_neighs=6)
sq.gr.nhood_enrichment(adata, cluster_key='leiden')
sq.gr.spatial_autocorr(adata, mode='moran', n_perms=100)
# Plotsprint('Creating plots...')
sc.pl.spatial(adata, color='leiden', spot_size=1.5, save='_clusters.pdf')
sq.pl.nhood_enrichment(adata, cluster_key='leiden', save='_nhood.pdf')
# Save
adata.write(f'{output_dir}/spatial_analyzed.h5ad')
print(f'Results saved to {output_dir}/')
Common Errors
Symptom
Cause
Fix
Spot clusters mislabeled as cell types
Skipped deconvolution on multi-cell Visium spots
Deconvolve against an annotated scRNA reference; clusters = niches (spatial-deconvolution)
Nearly all imaging cells filtered; small cells lost
Applied scRNA QC floor (min_counts=500) to single-molecule data
Gated on raw Moran's I / read a composition marker as within-type
Gate SVGs on FDR; a top-Moran gene usually marks a spatially-clustered cell TYPE
References
Palla G, Spitzer H, Klein M, et al (2022) Squidpy: a scalable framework for spatial omics analysis. Nature Methods 19:171-178. DOI 10.1038/s41592-021-01358-2. (spatial neighbor graph / neighborhood enrichment / autocorrelation.)
Kleshchevnikov V, Shmatko A, Dann E, et al (2022) Cell2location maps fine-grained cell types in spatial transcriptomics. Nature Biotechnology 40:661-671. DOI 10.1038/s41587-021-01139-4. (deconvolution: spot != cell.)
Cable DM, Murray E, Zou LS, et al (2022) Robust decomposition of cell type mixtures in spatial transcriptomics (RCTD). Nature Biotechnology 40:517-526. DOI 10.1038/s41587-021-00830-w.
Petukhov V, Xu RJ, Soldatov RA, et al (2022) Cell segmentation in imaging-based spatial transcriptomics with Baysor. Nature Biotechnology 40:345-354. DOI 10.1038/s41587-021-01044-w. (segmentation as the dominant imaging error source.)
