Spatial Analysis

Overview

This skill guides complete spatial transcriptomics analysis from raw data to biological insights. The fundamental question: What is the spatial organization of this tissue?

Workflow Decision Tree

START: User provides spatial data
    │
    ├─ Q: What platform?
    │   ├─ Visium/Visium HD → spot-based, may have histology
    │   ├─ Xenium/MERFISH/CosMx → single-cell resolution, imaging-based
    │   ├─ Slide-seq → spot-based, no histology
    │   └─ Unknown → check adata.uns for platform info
    │
    ├─ Q: Single sample or multiple?
    │   ├─ Single → proceed to standard workflow
    │   └─ Multiple → include integration step (Harmony recommended)
    │
    └─ Execute workflow based on answers

Standard Workflow

Step 1: Data Loading and Understanding

1. Load data using load_data tool
2. Examine dataset profile:
   - n_cells, n_genes
   - Available annotations (adata.obs columns)
   - Spatial coordinates availability
   - Tissue image availability
3. Understand the biological context from user

Key questions to ask user:

• What tissue/organ is this?
• What is your biological question?
• Do you have a reference dataset for cell type annotation?

Step 2: Quality Control and Preprocessing

Platform-specific QC thresholds:

| Metric | Visium | Xenium/MERFISH | Slide-seq |
|--------|--------|----------------|-----------|
| min_genes | 200 | 50 | 100 |
| min_cells | 3 | 3 | 3 |
| max_mito% | 20% | 10% | 15% |
| HVG count | 2000-3000 | 500-1000 | 1500-2000 |

Use preprocess_data tool with appropriate parameters.

Step 3: Spatial Domain Identification

Method Selection Guide:

Scenario	Recommended Method	Reason
Visium with histology	SpaGCN	Uses image features
Visium without histology	STAGATE/Leiden	Graph-based
Single-cell resolution	GraphST	Handles high resolution
Quick exploration	Leiden/Louvain	Fast, interpretable
Need fine structure	Higher resolution (1.5-2.0)	More clusters
Need broad domains	Lower resolution (0.3-0.5)	Fewer clusters

Use identify_spatial_domains tool.

Step 4: Validation and Refinement

Checklist:

• Do spatial domains match histological regions (if available)?
• Are marker genes for each domain biologically meaningful?
• Is the number of domains reasonable for the tissue?

Use find_markers to identify domain-specific genes. Use visualize_data with plot_type="feature" to examine spatial patterns.

Step 5: Multi-Sample Integration (if applicable)

When analyzing multiple samples:

1. Load all samples individually
2. Use integrate_samples with method selection:
   - harmony: Fast, good for batch correction (recommended default)
   - scvi: Deep learning, better for complex batches
   - bbknn: K-nearest neighbor based
   - scanorama: Panoramic stitching approach

3. Re-run clustering on integrated data
4. Validate integration quality

Step 6: Visualization and Reporting

Essential visualizations:

1. Spatial scatter colored by clusters/domains
2. UMAP embedding for global structure
3. Marker gene expression patterns
4. Violin plots for key genes

Use visualize_data tool with appropriate plot_type.

Platform-Specific Guidance

Visium

• Resolution: ~55μm spots, ~1-10 cells per spot
• Typical workflow: Full analysis → deconvolution for cell composition
• Histology integration: Use SpaGCN or include image features

Xenium/MERFISH

• Resolution: Single-cell
• Typical workflow: Full analysis → direct cell type annotation
• Large datasets: May need subsampling or chunked processing

Slide-seq

• Resolution: ~10μm beads
• No histology: Rely purely on expression patterns
• Bead-specific QC important

Output Expectations

A complete spatial analysis should provide:

1. Spatial domains with biological interpretation
2. Marker genes for each domain
3. Visualizations showing spatial organization
4. Quality metrics confirming data validity

Downstream Analysis Pointers

After establishing spatial structure, users may want:

• Cell type composition → use /cell-composition skill
• Cell-cell communication → use /cell-interaction skill
• Developmental trajectories → use /cell-dynamics skill
• Publication figures → use /publication-ready skill