Spatial Domain Detection - Usage Guide

Overview

This skill identifies spatial domains -- contiguous tissue regions of homogeneous expression such as cortical layers, tumor core, or stromal bands -- by combining transcriptional similarity with spatial proximity. It draws the line between a domain (a region containing many cell types), a cell type (one cell's identity), and a niche (the local cell-type composition), then guides choosing a domain method by tissue geometry, tuning the spatial-weight knob to avoid over-smoothing, and choosing the number of domains k as a biological decision.

Prerequisites

pip install squidpy scanpy scikit-learn
# optional domain methods:
pip install banksy_py            # neighbor-augmented features
pip install STAGATE_pyG          # graph attention autoencoder (or the TF build)
# BayesSpace / BASS / Banksy are R/Bioconductor:
# BiocManager::install(c('BayesSpace', 'Banksy'))

Quick Start

Tell your AI agent what you want to do: - "Identify spatial domains in my Visium section" - "Segment my tissue into regions, not cell types" - "My domains came out as smooth blobs -- fix the over-smoothing" - "Pick the number of domains and show me the k+-1 sensitivity"

Example Prompts

Domain vs Niche vs Cell Type

"Is this a domain question or a niche question? I want to know which tissue region each spot belongs to."

"Separate the cortical layers in my DLPFC section into spatially coherent regions."

Method Choice by Geometry

"Which domain method fits continuous laminar cortex versus a high-resolution Xenium tumor section?"

"My tumor nests are scattered, not contiguous -- which method or parameter handles non-contiguous domains?"

Tuning the Spatial-Weight Knob

"My domains are salt-and-pepper -- raise the spatial weight."

"My regions merged into blobs and the boundaries disappeared -- lower the over-smoothing."

"Sweep the BANKSY lambda and show me where the boundaries are sharpest."

Choosing k

"Set the number of domains for my section and report sensitivity to k plus or minus one."

"Run BayesSpace with q=7 for cortex and import the labels."

What the Agent Will Do

1. Establish whether the question is a domain (region), a niche (co-occurring types), or a cell type, and route accordingly.
2. Build a spatial neighbor graph appropriate to the platform (hex grid for Visium, kNN/Delaunay for imaging).
3. Run an expression-only baseline to show the salt-and-pepper failure a domain method corrects.
4. Run a neighbor-augmented or graph-based domain method, exposing the spatial-weight knob.
5. Sweep the spatial weight and inspect boundaries against histology to find the over-smoothing edge.
6. Choose k biologically and report k+-1 sensitivity.
7. Name domains from markers (for ranking/labeling, not inference) and validate against known architecture.

Tips

• A domain is a REGION with many cell types; a niche is which types co-occur; a cell type is one cell. If the question is "which types co-occur," use neighborhood enrichment in spatial-statistics, not domain segmentation.
• The knob that decides the outcome is the spatial weight (BANKSY lambda, BayesSpace smoothing, SpaGCN histology weight, GNN graph radius), not the clustering algorithm. Too much erases boundaries into blobs; too little reverts to salt-and-pepper.
• k is a biological choice, not a statistic to optimize. Report it and show k+-1.
• Non-contiguous domains (scattered tumor nests) break most methods because the spatial prior assumes contiguity (Yuan 2024). Lower the spatial weight or switch to niche analysis.
• DLPFC is the standard benchmark but it is continuous and laminar, which flatters smoothing methods. Do not over-generalize benchmark rankings to non-laminar tissue.
• No universal winner: GNN methods (STAGATE, GraphST, SEDR) and BayesSpace are robust on continuous Visium; BANKSY/BASS/stLearn lead on high-resolution imaging.
• In multi-sample data, use a multi-sample-aware method (BASS, PRECAST, GraphST integration) or batch-integrate first, or domains will track samples.

Related Skills

• spatial-neighbors - Build and tune the spatial graph every domain method inherits
• spatial-statistics - Niche/neighborhood enrichment when the question is which cell types co-occur
• spatial-deconvolution - Per-spot cell-type composition, a different question from regions
• single-cell/clustering - Non-spatial clustering and the double-dipping caveat on marker tests