By Lociven · SpatiaBio · August 2026
Visium spots are not single cells. Each spot captures RNA from multiple cells, and without knowing which cell types contributed what, your spatial analysis is blurry. Cell2location solves this by deconvolving Visium spots using a scRNA-seq reference. This post explains how it works and how to run it.
The problem: spots contain multiple cells
A standard Visium v2 spot is 55 µm in diameter. In most human tissues, that area contains 2–15 cells depending on cell size and packing density. When you cluster spots and call marker genes, you're clustering mixtures — and the "spatial domains" you find may reflect the composition of those mixtures rather than true biological boundaries.
Cell type deconvolution maps each spot to a probability distribution over cell types: "this spot is 60% cancer epithelial, 25% macrophage, 15% fibroblast." That's what cell2location produces.
Why cell2location over other methods? RCTD and NNLS are faster but assume cell type proportions sum to 1 (full tissue coverage). Cell2location uses a hierarchical Bayesian model that accounts for differences in RNA capture efficiency between scRNA-seq and Visium — which matters a lot in practice.
What you need
- Visium AnnData — raw counts (not normalized), spatial coordinates
- scRNA-seq reference — same tissue type, annotated cell types. Does not need to be the same sample.
- GPU strongly recommended (CPU works but is slow for large datasets)
The scRNA-seq reference is the critical input. If your reference has poorly annotated cell types, or doesn't cover all cell types present in the tissue, the deconvolution will be wrong. More cell types in the reference is generally better, as long as they're well-annotated.
Step 1: Prepare the scRNA-seq reference
import cell2location
import scanpy as sc
import numpy as np
# Load scRNA-seq reference
adata_ref = sc.read_h5ad("scrna_reference.h5ad")
# Keep raw counts
adata_ref.X = adata_ref.layers["counts"].copy()
# Filter to cell types with enough cells
cell_type_counts = adata_ref.obs["cell_type"].value_counts()
keep = cell_type_counts[cell_type_counts >= 5].index
adata_ref = adata_ref[adata_ref.obs["cell_type"].isin(keep)].copy()
# Filter genes
sc.pp.filter_genes(adata_ref, min_cells=1)
print(f"Reference: {adata_ref.n_obs} cells, {adata_ref.n_vars} genes")
print(f"Cell types: {adata_ref.obs['cell_type'].nunique()}")
Step 2: Train the reference model
from cell2location.models import RegressionModel
# Set up the regression model on the reference
RegressionModel.setup_anndata(
adata=adata_ref,
batch_key="sample", # batch correction if multiple samples
labels_key="cell_type",
layer="counts"
)
mod = RegressionModel(adata_ref)
mod.train(max_epochs=250, use_gpu=True)
# Export the estimated expression signatures
adata_ref = mod.export_posterior(
adata_ref,
sample_kwargs={"num_samples": 1000, "batch_size": 2500, "use_gpu": True}
)
# Save the cell type signatures
inf_aver = adata_ref.varm["means_per_cluster_mu_fg"][
[f"means_per_cluster_mu_fg_{i}" for i in adata_ref.uns["mod"]["factor_names"]]
].copy()
inf_aver.columns = adata_ref.uns["mod"]["factor_names"]
inf_aver.iloc[0:5, 0:5]
Step 3: Run deconvolution on Visium
import squidpy as sq
# Load Visium data
adata_vis = sq.read.visium("path/to/visium/")
adata_vis.var_names_make_unique()
# Keep only genes present in both datasets
intersect = np.intersect1d(adata_vis.var_names, inf_aver.index)
adata_vis = adata_vis[:, intersect].copy()
inf_aver = inf_aver.loc[intersect, :]
# Set up cell2location model on Visium
cell2location.models.Cell2location.setup_anndata(
adata=adata_vis, layer="counts"
)
mod_vis = cell2location.models.Cell2location(
adata_vis,
cell_state_df=inf_aver,
N_cells_per_location=10, # expected cells per spot
detection_alpha=20
)
mod_vis.train(
max_epochs=30000,
batch_size=None,
train_size=1,
use_gpu=True
)
Step 4: Extract and visualize cell type abundances
# Export posterior
adata_vis = mod_vis.export_posterior(
adata_vis,
sample_kwargs={"num_samples": 1000, "batch_size": 10, "use_gpu": True}
)
# Add 5% quantile cell type abundances to obsm
adata_vis.obs[adata_vis.uns["mod"]["factor_names"]] = adata_vis.obsm["q05_cell_abundance_w_sf"]
# Plot spatial distribution of each cell type
import matplotlib.pyplot as plt
cell_types_to_plot = ["Cancer_epithelial", "CD8_T_cell", "Macrophage", "Fibroblast"]
fig, axes = plt.subplots(1, len(cell_types_to_plot), figsize=(16, 4))
for ax, ct in zip(axes, cell_types_to_plot):
sq.pl.spatial_scatter(
adata_vis, color=ct, ax=ax,
cmap="magma", size=1.5, title=ct
)
plt.tight_layout()
plt.savefig("cell2location_output.png", dpi=150, bbox_inches="tight")
How deconvolution changes what a spot means. Left: a 55 µm Visium spot physically contains several cells of different types, and sequencing returns one averaged profile for all of them. Right: cell2location returns a composition instead — the worked example from this post. Read the q05 column, the conservative 5th-percentile estimate, rather than the mean.
Interpreting the output
The key output column is q05_cell_abundance_w_sf — the 5th percentile of the posterior distribution of cell type abundance per spot. Using the 5th percentile (rather than the mean) gives conservative, high-confidence estimates.
Common patterns to look for:
- Cancer cells concentrated in the tumor core, stromal cells at the margin
- CD8+ T cells enriched at the invasive margin vs excluded from the tumor core (a key immunotherapy biomarker)
- Macrophage subtypes (M1/M2) in different spatial zones
- Spots with very low total abundance — these may be necrotic areas or technical artifacts
Common mistake: Setting N_cells_per_location too high or too low drastically affects results. Use histology or DAPI staining to estimate the actual cell density in your tissue before running the model.
Using cell2location output with Squidpy
Once you have cell type abundances per spot, you can feed them directly into Squidpy's neighborhood enrichment and co-occurrence functions — treating cell types as categorical annotations rather than gene expression clusters.
import squidpy as sq
# Assign dominant cell type per spot
adata_vis.obs["dominant_cell_type"] = (
adata_vis.obs[adata_vis.uns["mod"]["factor_names"]]
.idxmax(axis=1)
)
# Now run neighborhood enrichment on deconvolved cell types
sq.gr.spatial_neighbors(adata_vis, coord_type="grid", n_rings=2)
sq.gr.nhood_enrichment(adata_vis, cluster_key="dominant_cell_type")
sq.pl.nhood_enrichment(adata_vis, cluster_key="dominant_cell_type", method="ward")
Get the complete pack
SpatiaBio Pack 2 — Deconvolution & Spatial Pattern Detection
The full cell2location deconvolution notebook — real scRNA-seq reference mapping, tissue zone clustering, and spatial neighborhood enrichment between zones. Plus a SpatialDE2 SVG detection notebook.
Get it for $19 →From the NeoantigenLab sister blog
Applying spatial transcriptomics to tumor immunology?
NeoantigenLab covers the biology behind what you're profiling — neoantigen pipelines, TIL therapy, checkpoint inhibitors, and cancer immunotherapy for experimental researchers.
Visit NeoantigenLab →Tags: cell2location, cell type deconvolution, Visium, spatial transcriptomics, scRNA-seq, RCTD, Bayesian deconvolution, tumor microenvironment, Squidpy
No comments:
Post a Comment