Differential Analysis

Overview

This skill answers: What genes/features differ between groups in this spatial data?

Three main comparison types:

1. Marker genes: What defines each cluster/cell type?
2. Condition comparison: How does treatment/disease change expression?
3. Regional comparison: What differs between spatial domains?

Decision Tree: Which Comparison?

Q: What are you comparing?
│
├─ Clusters or cell types
│   └─ Marker Finding
│       ├─ Wilcoxon rank-sum - Default, robust
│       ├─ t-test - Parametric, fast
│       └─ Logistic regression - Handles confounders
│
├─ Experimental conditions (treated vs control)
│   └─ Condition Comparison
│       ├─ Pseudobulk + DESeq2 - Gold standard for replicates
│       ├─ MAST - Single-cell aware
│       └─ Mixed models - Complex designs
│
└─ Spatial domains or regions
    └─ Regional Comparison
        ├─ Same as marker finding
        └─ Consider spatial autocorrelation

Marker Gene Finding

Purpose

Identify genes that distinguish one group from all others (1 vs rest) or from a specific group (pairwise).

Method Selection

Workflow

Step 1: Define Groups

Ensure grouping column exists in adata.obs:

• Clusters: leiden, louvain
• Cell types: cell_type, annotation
• Domains: spatial_domain

Step 2: Run Marker Finding

Use find_markers tool with:

• groupby: Column name defining groups
• method: "wilcoxon", "t-test", or "logreg"
• n_genes: Number of top markers to return per group

Step 3: Interpret Results

Key output columns:

• names: Gene name
• scores: Test statistic
• pvals_adj: Adjusted p-value (FDR corrected)
• logfoldchanges: Log2 fold change
• pct_nz_group: Percent expressing in group
• pct_nz_reference: Percent expressing in other groups

Quality markers criteria:

Strong markers typically have:
- logFC > 1 (2-fold change) or < -1
- padj < 0.05
- pct_nz_group > 0.25 (expressed in >25% of group)
- pct_nz_group >> pct_nz_reference (specificity)

Condition Comparison

Purpose

Compare expression between experimental conditions (e.g., disease vs healthy, treated vs control).

Considerations

Replicates matter:

• With replicates: Use pseudobulk + DESeq2/edgeR
• Without replicates: Treat as exploratory only

Confounders:

• Batch effects: Include batch as covariate
• Cell type composition: Control or stratify

Workflow

Step 1: Set Up Comparison

Identify:

• Condition column in adata.obs
• Reference group (baseline/control)
• Confounding variables to adjust

Step 2: Run Comparison

Use compare_conditions tool with:

• condition_key: Column with conditions
• reference: Reference/control group name
• method: "pseudobulk", "mast", or "wilcoxon"
• covariates: Optional confounders to adjust

Step 3: Interpret Results

Volcano plot interpretation:

• X-axis: Log2 fold change (effect size)
• Y-axis: -log10(p-value) (significance)
• Upper right: Upregulated in condition
• Upper left: Downregulated in condition

Thresholds:

Significant DEGs typically:
- |logFC| > 0.5 (at least 1.4-fold change)
- padj < 0.05
- For stricter: |logFC| > 1, padj < 0.01

Regional/Spatial Domain Comparison

Purpose

Compare expression between different spatial regions or domains.

Special Considerations

Spatial autocorrelation:

• Nearby spots are not independent
• Standard tests may be anti-conservative
• Consider spatial-aware methods or pseudobulk

Workflow

Step 1: Define Regions

Regions can come from:

• Spatial domain identification
• Manual annotation
• Anatomical boundaries

Step 2: Run Comparison

Use find_markers with regions as groups:

• Pairwise: Compare specific domains
• 1 vs rest: Find domain-specific genes

Step 3: Validate Spatially

Visualize top DEGs on spatial map:

• Confirm pattern matches expected region
• Check for artifacts at boundaries

Visualization

Essential Plots

1. Volcano plot: Overview of all comparisons
2. Heatmap: Top markers across groups
3. Violin plots: Distribution of key genes
4. Spatial maps: Expression in tissue context

Use visualize_data with:

• plot_type="volcano" for volcano plots
• plot_type="heatmap" for expression heatmaps
• plot_type="violin" for distributions

Multiple Testing Correction

Why It Matters

Testing thousands of genes inflates false positives:

• 20,000 genes × 0.05 = 1,000 false positives expected

Correction Methods

Always use adjusted p-values for final results!

Common Analysis Scenarios

Scenario 1: Find Cell Type Markers

Goal: What genes define each annotated cell type?
1. find_markers with groupby="cell_type"
2. Filter: padj < 0.05, logFC > 1, pct_nz > 0.25
3. Validate known markers present
4. Visualize top 3-5 per type

Scenario 2: Disease vs Healthy

Goal: What changes in disease state?
1. compare_conditions with condition_key="disease_state"
2. Set reference="healthy"
3. Add covariates if batch effects present
4. Volcano plot to visualize
5. Pathway analysis on significant DEGs

Scenario 3: Tumor Margin Analysis

Goal: What genes differ at tumor-stroma boundary?
1. Define regions: "tumor_core", "margin", "stroma"
2. find_markers with groupby="region"
3. Focus on margin-specific genes
4. Cross-reference with communication analysis

Troubleshooting

Output Summary

Successful differential analysis provides:

• DEG list: Ranked genes with statistics
• Effect sizes: Log fold changes
• Significance: Adjusted p-values
• Visualizations: Volcano, heatmap, spatial maps
• Downstream input: Gene lists for enrichment analysis