Bio Workflows Imc Pipeline

# generate the panel template; edit the keep column before extracting
steinbock preprocess imc panel

# extract per-channel TIFFs (keep-filtered, panel-ordered) with hot-pixel removal
# (--hpf is a signed 8-neighbor difference; 50 is a count, tune to dynamic range)
steinbock preprocess imc images --hpf 50

# channel spillover is compensated with NNLS (CATALYST/cytomapper, R) on the pixel images
# BEFORE segmentation when spatial analysis is the endpoint -- see data-preprocessing

# Mesmer/DeepCell whole-cell (nuclear-first); membrane channels aggregated via the panel column.
# --pixelsize is steinbock's CLI flag for the acquisition resolution (it wraps Mesmer's image_mpp);
# steinbock defaults it to 1.0 um for IMC, so pass the true value explicitly rather than relying on it.
steinbock segment deepcell --pixelsize 1.0 --minmax -o masks

# Alternative: Cellpose container (Cellpose 4+ default model cpsam; channel order reversed vs native)
steinbock segment cellpose --minmax -o masks

# Extract per-cell MEAN intensities (mean is the default and the right phenotyping aggregator;
# sum confounds cell size with expression)
steinbock measure intensities -o intensities

# Measure cell properties (area, centroid, eccentricity)
steinbock measure regionprops -o regionprops

# Build the spatial neighbor graph (expansion within a max distance; match the graph to the
# biological claim -- contact vs proximity -- in spatial-analysis)
steinbock measure neighbors --type expansion --dmax 15 -o neighbors

import pandas as pd
import numpy as np
import anndata as ad
import scanpy as sc
import squidpy as sq
from pathlib import Path

# === 1. LOAD DATA ===
data_dir = Path('steinbock_output')

intensities = pd.read_csv(data_dir / 'intensities.csv', index_col=0)
regionprops = pd.read_csv(data_dir / 'regionprops.csv', index_col=0)
neighbors = pd.read_csv(data_dir / 'neighbors.csv')

print(f'Loaded {len(intensities)} cells')

# === 2. CREATE ANNDATA ===
adata = ad.AnnData(X=intensities.values, obs=regionprops, var=pd.DataFrame(index=intensities.columns))
adata.obs['image_id'] = pd.Categorical([idx.rsplit('_', 1)[0] for idx in intensities.index])   # strip only the trailing cell index: rsplit keeps Patient1_ROI002 distinct from Patient1_ROI001. squidpy library_key requires a categorical, not object/string
adata.obs['cell_id'] = intensities.index

# Add spatial coordinates (skimage regionprops_table names them centroid-0 (y) / centroid-1 (x))
adata.obsm['spatial'] = regionprops[['centroid-0', 'centroid-1']].values

# === 3. PREPROCESSING ===
# Arcsinh transform: cofactor 1 for IMC single-cell means (Hunter 2024), NOT the
# suspension-CyTOF cofactor 5, which over-compresses IMC's lower-count means
adata.layers['counts'] = adata.X.copy()
adata.X = np.arcsinh(adata.X / 1)

# Scale for clustering
sc.pp.scale(adata, max_value=10)
adata.raw = adata.copy()

# === 4. DIMENSIONALITY REDUCTION ===
sc.pp.pca(adata, n_comps=20)
sc.pp.neighbors(adata, n_neighbors=15)
sc.tl.umap(adata)

# === 5. CLUSTERING ===
sc.tl.leiden(adata, resolution=0.8)
print(f'Found {adata.obs["leiden"].nunique()} clusters')

# === 6. PHENOTYPING ===
# Marker expression per cluster
sc.tl.rank_genes_groups(adata, 'leiden', method='wilcoxon')
marker_genes = sc.get.rank_genes_groups_df(adata, group=None)

# Annotate clusters based on markers
cluster_annotations = {
    '0': 'T cells',
    '1': 'Macrophages',
    '2': 'Tumor',
    '3': 'B cells',
    '4': 'Stromal'
}
adata.obs['cell_type'] = adata.obs['leiden'].map(cluster_annotations)

# === 7. SPATIAL ANALYSIS ===
# Build spatial graph PER IMAGE (library_key), else Delaunay fabricates edges across ROIs
sq.gr.spatial_neighbors(adata, coord_type='generic', delaunay=True, library_key='image_id')

# Neighborhood enrichment
sq.gr.nhood_enrichment(adata, cluster_key='cell_type')

# Co-occurrence analysis
sq.gr.co_occurrence(adata, cluster_key='cell_type')

# Ripley's statistics
sq.gr.ripley(adata, cluster_key='cell_type', mode='L')

# === 8. VISUALIZATION ===
import matplotlib.pyplot as plt

# UMAP by cell type
fig, axes = plt.subplots(1, 2, figsize=(14, 5))
sc.pl.umap(adata, color='cell_type', ax=axes[0], show=False)
sc.pl.umap(adata, color='leiden', ax=axes[1], show=False)
plt.savefig('umap_celltypes.png', dpi=150, bbox_inches='tight')

# Spatial plot. Pick the image dynamically: image_id is derived from the cell index, so a hardcoded
# literal selects zero cells and spatial_scatter errors on the empty subset.
fig, ax = plt.subplots(figsize=(10, 10))
first_image = adata.obs['image_id'].iloc[0]
sq.pl.spatial_scatter(adata[adata.obs['image_id'] == first_image],
                      color='cell_type', shape=None, size=10, ax=ax)
plt.savefig('spatial_celltypes.png', dpi=150, bbox_inches='tight')

# Neighborhood enrichment heatmap
sq.pl.nhood_enrichment(adata, cluster_key='cell_type')
plt.savefig('neighborhood_enrichment.png', dpi=150, bbox_inches='tight')

# === 9. DIFFERENTIAL ANALYSIS (patient is the unit, NOT the cell) ===
import statsmodels.formula.api as smf

# aggregate to per-image proportions, then test across PATIENTS -- a cell-level or per-image
# test over correlated cells is pseudoreplication and reports p~0 for trivial effects.
# obs must carry patient and condition columns; see differential-analysis for scCODA
# (compositional) and the spatial differential path.
counts = adata.obs.groupby(['patient', 'condition', 'image_id', 'cell_type'], observed=True).size().unstack(fill_value=0)   # observed=True: image_id is categorical; the default expands the full cartesian product into all-zero phantom rows -> NaN proportions
image_prop = counts.div(counts.sum(axis=1), axis=0).reset_index()
target = 'Tumor'   # an actual cell_type column from cluster_annotations above (single-word for the formula)
res = smf.mixedlm(f'{target} ~ condition', image_prop, groups=image_prop['patient']).fit()  # patient random effect
print(res.summary())

adata.write('imc_analysis.h5ad')
print('Analysis complete!')

library(imcRtools)
library(cytomapper)
library(CATALYST)

# Read steinbock output
spe <- read_steinbock('steinbock_output/')

# Transform (cofactor 1 for IMC single-cell means, not 5)
assay(spe, 'exprs') <- asinh(counts(spe) / 1)

# Cluster (CATALYST runDR takes assay=; cluster() always uses the 'exprs' assay, no assay arg)
spe <- runDR(spe, features = rownames(spe), assay = 'exprs', dr = 'UMAP')
spe <- cluster(spe, features = rownames(spe), xdim = 10, ydim = 10, maxK = 20)

# Spatial analysis. buildSpatialGraph names the colPair '<type>_interaction_graph';
# aggregateNeighbors counts a label via aggregate_by='metadata' + count_by=.
spe <- buildSpatialGraph(spe, img_id = 'sample_id', type = 'expansion', threshold = 20)
spe <- aggregateNeighbors(spe, colPairName = 'expansion_interaction_graph',
                          aggregate_by = 'metadata', count_by = 'cluster_id')

# Spatial context
spe <- detectCommunity(spe, colPairName = 'expansion_interaction_graph',
                       size_threshold = 10, group_by = 'sample_id')

# Plot (img_id is the colData COLUMN used to facet; read_steinbock names it 'sample_id', not 'image_id')
plotSpatial(spe, img_id = 'sample_id', node_color_by = 'cluster_id')

# Use batch-aware clustering
import scvi

scvi.model.SCVI.setup_anndata(adata, batch_key='image_id')
model = scvi.model.SCVI(adata)
model.train()
adata.obsm['X_scvi'] = model.get_latent_representation()
sc.pp.neighbors(adata, use_rep='X_scvi')

# Spatial cell-cell co-location around tumor (per-image, then aggregate to patient).
# Note: sq.gr.ligrec keys ligand-receptor pairs on gene symbols from OmniPath, so it is
# usually empty on a ~40-marker antibody panel -- prefer neighborhood enrichment for IMC.
sq.gr.nhood_enrichment(adata, cluster_key='cell_type')   # see spatial-analysis for the null caveat