Spatial Analysis - Usage Guide

Overview

Tests cell-cell co-location, discovers tissue niches, and quantifies spatial structure from segmented IMC/MIBI cell tables. The point that shapes every result here: a spatial interaction or niche is a hypothesis test against a null model, and the null chosen (label-shuffle vs CSR vs density-conditioned vs patient-level) silently decides whether the result is real biology, a density gradient, a segmentation artifact, or upstream clustering -- so the skill forces two questions onto every analysis: against which null, and at which unit.

Prerequisites

pip install squidpy scanpy anndata scikit-learn numpy
# R for inhomogeneous Ripley's K and imcRtools graphs (optional)
# install.packages('spatstat'); BiocManager::install('imcRtools')

Quick Start

Tell your AI agent what you want to do: - "Test whether CD8 T cells co-locate with tumor cells, with the right null" - "Build a contact graph and prune the long edges" - "Discover cellular neighborhoods and check they are stable to window size" - "Tell me if this interaction is real or just a density artifact" - "Compute per-image neighborhood enrichment to aggregate across patients"

Example Prompts

Neighborhood enrichment

"Run neighborhood enrichment between all cell types per image. Tell me which null this uses and what it cannot control for."

Density confound

"My CD8 and Treg cells both live in the tumor margin. Is their 'interaction' real, or are they just sharing a compartment? Cross-check with inhomogeneous Ripley's K."

Niche discovery

"Find cellular neighborhoods in my cohort and show me whether the niches survive changing the window from 10 to 30 neighbors."

Graph choice

"I want to test contact-dependent (juxtacrine) interactions. Which graph should I build and how do I avoid edges across empty tissue?"

What the Agent Will Do

Build a spatial graph matching the biological claim -- pruned Delaunay for contact, fixed radius for paracrine proximity -- per image.
Run neighborhood enrichment per image, naming the null and treating the z-score as a per-image summary (never thresholded across images of different size).
Cross-check density-driven hits with inhomogeneous cross-K.
Discover cellular neighborhoods, sweeping the window size and reporting stability.
Hand cross-condition comparison (does a niche/interaction differ between groups) to differential-analysis at the patient unit, with FDR across pairs.

Tips

The label-shuffle null controls abundance but not local density -- two types sharing a compartment will look "interacting".
The squidpy z-score is unbounded and graph-degree-dependent; it is not comparable across images of different cell counts.
Delaunay invents long edges across lumen/necrosis; prune by a biological max distance before counting.
kNN graphs mix contact and proximity by density; state the micron scale of any "interaction".
A cellular neighborhood has no built-in significance test; it is exploratory until compared across patients.
Verify that niche-defining markers are not one cell's signal bleeding into a neighbor (spillover).

phenotyping - cell-type labels are the input to every spatial test
differential-analysis - testing whether interactions/niches differ between conditions at the patient level
cell-segmentation - over-segmentation and lateral spillover create fake niches
data-preprocessing - uncompensated spillover manufactures false cell-cell interactions
spatial-transcriptomics/spatial-statistics - shared squidpy neighborhood and autocorrelation methods