How to Annotate Cell Types in scRNA-seq: Complete Guide

Prepare Your scRNA-seq Data

Before annotating cell types, you need to extract marker genes from your clustered scRNA-seq data.

Using Seurat (R):

# Find markers for all clusters
markers <- FindAllMarkers(seurat_object, 
                         only.pos = TRUE, 
                         min.pct = 0.25, 
                         logfc.threshold = 0.25)

# Export top markers
top_markers <- markers %>% 
  group_by(cluster) %>% 
  top_n(n = 50, wt = avg_log2FC)

write.csv(top_markers, "markers_for_annotation.csv")

Using Scanpy (Python):

# Find marker genes
sc.tl.rank_genes_groups(adata, 'clusters', method='wilcoxon')

# Export to DataFrame
markers_df = pd.DataFrame(adata.uns['rank_genes_groups']['names']).head(50)
markers_df.to_csv('markers_for_annotation.csv')

Choose Your Annotation Method

Traditional Methods:

• Manual Annotation: Time-consuming but accurate for experts
• Reference-Based: SingleR, Seurat mapping (requires good references)
• Marker-Based: CellMarker, PanglaoDB (limited coverage)

AI-Powered Method (Recommended):

Multi-model consensus annotation using mLLMCelltype provides:

• Higher accuracy through consensus
• No reference requirements
• Handles novel cell types
• Automated workflow

Upload Your Marker Genes

Navigate to mLLMCelltype and upload your marker genes file.

Supported Formats:

• CSV with clusters as columns
• TSV with marker gene lists
• Excel files from common tools

Configure Annotation Parameters

Essential Settings:

• Species: Human, Mouse, or custom
• Tissue Type: PBMC, Brain, Liver, etc.
• AI Models: Select multiple for consensus (GPT-4, Claude, Gemini)

Advanced Options:

• Consensus Threshold: 0.6 (default) - increase for stricter consensus
• Entropy Threshold: 1.2 (default) - lower for higher confidence
• Discussion Rounds: 3 (default) - more rounds for difficult cases

Run Automated Annotation

Click "Start Analysis" to begin the multi-model annotation process.

What Happens Next:

1. Each AI model analyzes your marker genes independently
2. Models compare their annotations
3. If disagreement occurs, models discuss to reach consensus
4. Final annotations are determined with confidence scores

Interpret and Download Results

Understanding Your Results:

• Cell Type: The consensus annotation for each cluster
• Consensus Score: Agreement level between models (0-1)
• Entropy: Uncertainty measure (lower is better)
• Individual Predictions: Each model's annotation

Next Steps:

1. Download results as CSV
2. Import back to Seurat/Scanpy
3. Visualize on UMAP/tSNE
4. Validate key populations

Best Practices for Cell Type Annotation

Data Quality:

• Ensure proper QC filtering before clustering
• Remove doublets and low-quality cells
• Use appropriate clustering resolution

Marker Gene Selection:

• Use statistically significant markers (adjusted p-value < 0.05)
• Include both positive and negative markers
• Consider fold-change thresholds

Validation:

• Check known marker expression
• Compare with published datasets
• Perform downstream functional analysis

Troubleshooting Common Issues

Low Confidence Annotations:

If you get low confidence scores:

• Check if clusters are well-separated
• Ensure sufficient marker genes
• Consider re-clustering at different resolutions

Unexpected Cell Types:

If annotations seem incorrect:

• Verify tissue type is correctly specified
• Check for batch effects
• Review marker gene quality