Spatial Phenotyping

Spatial Phenotyping computes the cellular neighborhood composition for each cell:

1. Cell positions — Centroids from the coded image provide each cell's (x, y) location.
2. Cell types — Phenotype assignments from classification provide each cell's type label.
3. Neighborhood search — For each cell, all other cells within the specified radius are identified.
4. Composition computation — The proportions of each cell type within the neighborhood are calculated. Cell #47 might have 60% tumor, 25% T cell, 10% stromal, 5% B cell within its 50 µm radius.
5. Spatial phenotype assignment — Based on the neighborhood composition, cells are assigned spatial phenotype labels (e.g., "immune-enriched tumor border", "immune-desert tumor core").

Markov random fields and spatial context: The MIT Statistical Models chapter describes Markov random fields (MRFs) — probabilistic models where each variable's distribution depends only on its neighbors. In tissue biology, cells exhibit exactly this Markov-like behavior: a cell's phenotype and function are influenced by its local neighborhood but largely independent of distant cells. Spatial phenotyping quantifies this neighborhood influence by measuring what types of cells share the local context.

Cellular automata as conceptual model: Gardner's Game of Life shows how simple local rules (counting neighbors of each type) produce complex global patterns — still lifes, oscillators, gliders. Tissue architecture similarly emerges from local cell-cell interactions. Spatial phenotyping measures the "rules" — the neighborhood compositions — that generate the observed tissue patterns. A region where tumor cells have many immune neighbors is biologically different from one where they don't, even if the global cell counts are identical.

Spatial point processes: The collection of cell positions forms a spatial point process. A completely random (Poisson) process would show no spatial structure — any cell type would be equally likely at any position. Real tissue shows marked departures: immune cells cluster at the tumor boundary, tumor cells form coherent nests, stromal cells fill the spaces between. Spatial phenotyping characterizes these departures from randomness at the individual-cell level.

The spatial scale matters: Different biological processes operate at different spatial scales. Cell-cell signaling (cytokines, direct contact) operates at 10-50 µm. Microenvironment zones (tumor nests, immune aggregates) operate at 50-200 µm. Tissue architecture (invasive margin, tumor center) operates at 200+ µm. The choice of neighborhood radius determines which scale of spatial organization is captured. Multi-scale analysis — running spatial phenotyping at several radii — can reveal patterns invisible at any single scale.

Characterizing the tumor-immune interface in a PD-L1 study:

1. Classify cells: CK+ tumor, CD3+CD8+ cytotoxic T, CD3+CD8− helper T, CD20+ B, other
2. Spatial Phenotyping with 50 µm radius
3. Identify spatial phenotypes:
   • "Immune hot" tumor cells: CK+ cells with >30% CD8+ neighbors
   • "Immune cold" tumor cells: CK+ cells with <5% immune neighbors
   • "Engaged" T cells: CD8+ cells with >50% tumor neighbors
   • "Bystander" T cells: CD8+ cells with <10% tumor neighbors
4. Result: Spatial phenotype map reveals the tumor-immune interface architecture, identifying where immune cells are actively engaging tumor cells versus where they're excluded