Spatial Transcriptomics Pipeline - Usage Guide

Overview

This workflow analyzes spatial transcriptomics data (Visium, Xenium) from raw data to spatial domains and visualizations using Squidpy and Scanpy. It branches first on the platform-class fork: sequencing/capture data (Visium) are spot mixtures that are QC'd, clustered into regions, and deconvolved for composition; imaging data (Xenium) are single cells that need low-count QC floors and segmentation. Composition estimation (spatial-deconvolution) and cell-cell communication (spatial-communication) are handed off to those skills rather than inlined.

Prerequisites

pip install squidpy scanpy matplotlib

Quick Start

Tell your AI agent what you want to do: - "Analyze my Visium spatial transcriptomics data" - "Find spatially variable genes in my tissue" - "Identify spatial domains in my sample"

Example Prompts

Loading and QC

"Load my Space Ranger output"

"Show QC metrics on the tissue image"

Analysis

"Find spatially variable genes"

"Run neighborhood enrichment analysis"

"Detect spatial domains"

Visualization

"Plot gene expression on the tissue"

"Show clusters overlaid on the image"

Input Requirements

What the Workflow Does

1. Load Data - Read spatial data with images (squidpy for Visium; spatialdata_io for Xenium, which has no squidpy reader)
2. QC - Filter spots with platform-appropriate floors (Visium floors delete real imaging cells, so branch on the fork)
3. Clustering - Cluster spots into regions/niches, not cell types (a Visium spot is a 1-10-cell mixture)
4. Spatial Analysis - Neighbor graph on the hex lattice (coord_type grid), Moran's I gated on FDR, neighborhood enrichment read against its weak null
5. Domains - Spatial domain detection (a domain is a region, distinct from a cell type and a niche)
6. Visualization - Plots on tissue with the correct spot size and coordinate frame

Tips

• Platform fork: decide imaging vs sequencing first; it sets QC floors and whether you deconvolve or segment
• QC floors: Visium UMI/gene floors are tissue-dependent starting points; never apply a 500-count floor to imaging data
• SVGs: gate top Moran's I on FDR, and separate composition-driven SVGs (cell-type markers) from within-type regulation
• nhood enrichment: a positive z is not a specific interaction; the global null is confounded by abundance and compartments
• Deconvolution and communication: run them through spatial-deconvolution and spatial-communication, which carry the reference and spillover caveats