Bio Imaging Mass Cytometry Spatial Analysis

Version Compatibility

Reference examples tested with: squidpy 1.3+, scanpy 1.10+, anndata 0.10+, numpy 1.26+, imcRtools 1.8+ (R), spatstat 3.0+ (R)

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.

Notes specific to this skill: squidpy's nhood_enrichment z-score is unbounded and scales with graph degree, so it is NOT comparable across images of different cell counts. squidpy does not implement Ripley's K (its ripley and co_occurrence are loose analogs); the inhomogeneous K that conditions on tissue density lives in R spatstat::Kinhom/Kcross.inhom. Build the graph with gr.spatial_neighbors(coord_type='generic') before any test.

Spatial Analysis for IMC

"Analyze spatial cell interactions in my IMC data" -> Build a neighbor graph, then test cell-type co-location against an explicitly chosen null, or discover recurrent niches.

• Python: squidpy.gr.spatial_neighbors, squidpy.gr.nhood_enrichment, squidpy.gr.co_occurrence
• R: imcRtools::buildSpatialGraph, spatstat::Kcross.inhom

The Single Most Important Modern Insight -- a spatial interaction is a hypothesis test, and the null silently decides what is measured

A "spatial interaction" or "niche" is not an observation; it is a test against a null model, and which null is chosen (label-shuffle vs CSR vs density-conditioned vs patient-level) decides whether the result is real biology, a density gradient, a segmentation artifact, or upstream clustering. The label-permutation null (histoCAT, squidpy nhood_enrichment) holds cell positions fixed and shuffles identities: it DOES control global abundance (a rare type cannot show spurious enrichment just from being rare) but it does NOT control local density or tissue architecture -- so two cell types that merely share an anatomical compartment (both enriched in a follicle or at an invasive margin) score as strongly "interacting" with no direct affinity. This density confound is the dominant false-positive engine, and a finding significant under CSR but null under inhomogeneous Ripley's K is a density artifact, not an interaction. Two further facts compound it: the nhood_enrichment z-score is unbounded and graph-degree-dependent, so a z of 30 in a 50k-cell image and a z of 8 in a 5k-cell image are not on the same scale (never threshold a fixed z across heterogeneous images); and the replicate is the patient, not the cell or the image, so a per-cell test over tens of thousands of cells reports p~0 for trivial effects (pseudoreplication). The skill forces two questions onto every analysis: against which null, and at which unit.

Spatial Methods Taxonomy

Decision Tree by Scenario

Build the Spatial Graph (contact vs proximity)

Goal: Construct the graph whose definition matches the biological claim.

Approach: Delaunay approximates physical adjacency (contact/juxtacrine) but invents long edges across lumen/necrosis, so prune by a max distance; fixed radius gives true proximity (paracrine) at a stated micron scale; kNN silently mixes contact and proximity because fixed k spans microns in dense regions and hundreds of microns in sparse ones. Build the graph per image.

import squidpy as sq
import numpy as np

# contact graph: Delaunay, then prune edges longer than a biological max distance (um)
sq.gr.spatial_neighbors(adata, coord_type='generic', delaunay=True)
dist = adata.obsp['spatial_distances']
keep = dist.copy(); keep.data[keep.data > 30] = 0; keep.eliminate_zeros()   # cap at ~30 um
adata.obsp['spatial_connectivities'] = (keep > 0).astype(float)

# OR proximity graph: fixed radius at a justified micron scale (paracrine range)
# sq.gr.spatial_neighbors(adata, coord_type='generic', radius=30.0)

Neighborhood Enrichment with a Named Null

Goal: Test co-location per image with the abundance-aware permutation null, knowing its blind spot.

Approach: Run nhood_enrichment per image (so the shuffle is within-image), keep the z as a per-image summary, and never threshold a fixed z across images of different size. Cross-check density-driven hits against inhomogeneous K.

per_image_z = {}
for img_id, idx in adata.obs.groupby('image_id').groups.items():
    sub = adata[idx].copy()
    sq.gr.spatial_neighbors(sub, coord_type='generic', delaunay=True)
    sq.gr.nhood_enrichment(sub, cluster_key='cell_type', seed=0)
    per_image_z[img_id] = sub.uns['cell_type_nhood_enrichment']['zscore']
# aggregate these per-image summaries to the PATIENT unit in differential-analysis, not here

Cellular Neighborhoods (niche discovery)

Goal: Find recurrent local cell-type compositions, treating them as exploratory.

Approach: Per cell, summarize the composition of its window of neighbors, then cluster the windows. The window size IS the spatial scale and is almost always unjustified, so sweep it and report that the biological conclusion survives k in {10, 20, 30}. A CN has no built-in significance test; significance enters only as a cross-condition comparison (differential-analysis).

from sklearn.cluster import KMeans
import pandas as pd

def cellular_neighborhoods(adata, k_window=20, n_cn=10):
    sq.gr.spatial_neighbors(adata, coord_type='generic', n_neighs=k_window)   # window = k neighbors
    conn = adata.obsp['spatial_connectivities']
    onehot = pd.get_dummies(adata.obs['cell_type']).values
    comp = (conn @ onehot)                       # neighbor composition per cell
    comp = comp / comp.sum(axis=1, keepdims=True).clip(min=1)
    return KMeans(n_clusters=n_cn, random_state=0).fit_predict(comp)

adata.obs['CN'] = cellular_neighborhoods(adata, k_window=20)   # sweep k_window to check stability

Per-Method Failure Modes

nhood_enrichment -- density false positive

Trigger: reporting a z-score as a "significant interaction." Mechanism: the label-shuffle null absorbs abundance but not local density, so co-compartmentalized types score as interacting. Symptom: every type pair sharing a region looks attracted. Fix: name the null; cross-check with inhomogeneous cross-K; or shuffle within a compartment, not the whole image.