jaechang-hits/single-cell-annotation

Single Cell RNA-seq Cell Type Annotation

---

Metadata

Short Description: Best practices for annotating cell types in single-cell RNA-seq data using marker-based, automated, and reference-based approaches.

Authors: Distilled from "Single-cell best practices" by Luecken, M.D. et al.

Affiliations: Helmholtz Munich, Wellcome Sanger Institute, Harvard Medical School, and contributors

Version: 1.0

Last Updated: January 2025

License: CC BY 4.0

Commercial Use: ✅ Allowed

Source: https://www.sc-best-practices.org/cellular_structure/annotation.html

Citation: Luecken, M.D., Theis, F.J. et al. (2023). Current best practices in single-cell RNA-seq analysis: a tutorial. Molecular Systems Biology.

---

Overview

Cell type annotation is the process of assigning cell type labels to clusters or individual cells in single-cell RNA-seq data. This guide covers three main approaches and their practical implementation.

Key Concepts

Cell Type vs. Cell State

A cell type is a stable identity defined by a developmental trajectory and core marker gene program (e.g., CD4+ T cell, hepatocyte). A cell state is a transient condition (activated, cycling, stressed) overlaid on a cell type. Annotation should target cell types first; states are attributes that may further subdivide a type but should not be conflated with type identity.

Marker Genes and Marker Panels

Marker genes are genes whose expression is enriched in a specific cell type relative to other cells in the same tissue context. Reliable annotation uses panels of multiple markers (typically 3-5 per type) rather than a single gene, because expression is noisy in droplet-based scRNA-seq and many markers are shared across related types. Markers come in two flavors: canonical (literature-derived, e.g., CD3D for T cells) and data-derived (from differential expression on the dataset).

Reference Atlases and Label Transfer

A reference atlas is a previously annotated dataset (e.g., Human Cell Atlas, Tabula Sapiens) used to project labels onto a new "query" dataset. Label transfer methods (scArches, scANVI, Azimuth, SingleR) align query cells into the reference latent space and assign the nearest neighbor's label. Quality of transfer depends on tissue match, technology match (e.g., 10x v3 vs. Smart-seq2), and species match.

Decision Framework

Use this tree to choose an annotation approach:

                Do you have a well-characterized tissue
                with a high-quality reference atlas?
                            │
              ┌─────────────┴─────────────┐
              │                           │
             YES                          NO
              │                           │
              ▼                           ▼
   Is this a standard tissue       Are you studying
   (PBMC, lung, gut) with a       novel cell types or
   pre-trained classifier?         exploratory data?
              │                           │
        ┌─────┴─────┐               ┌─────┴─────┐
        │           │               │           │
       YES          NO             YES          NO
        │           │               │           │
        ▼           ▼               ▼           ▼
   Automated    Reference-     Manual marker  Manual +
   (CellTypist) based          based          automated
                (scArches,     (Scanpy,       cross-check
                 Azimuth,      Seurat)
                 SingleR)

Decision Table

| Scenario | Approach | Primary Tool | Validation | |----------|----------|--------------|------------| | Standard human PBMC, large dataset (>100k cells) | Automated | CellTypist | Spot-check with manual markers | | Well-characterized tissue (lung, kidney, brain) | Reference-based label transfer | scArches / Azimuth | Marker consistency on top clusters | | Novel/rare tissue, no good reference | Manual marker-based | Scanpy / Seurat | Hierarchical, broad-to-fine | | Cross-species (e.g., zebrafish) | Manual markers + ortholog mapping | Scanpy + custom panel | Compare to closest reference species | | Developmental / continuous trajectory | Reference-based with state-aware model | scANVI / scArches | Trajectory coherence + markers | | Disease tissue with known perturbation | Manual + automated cross-check | CellTypist + Scanpy | Confirm disease-specific states separately |

Three Annotation Approaches

1. Manual Marker-Based Annotation

Identify cell types by examining expression of known marker genes in each cluster.

Tools: Scanpy, Seurat Best for: Small datasets, novel cell types, high confidence needs

2. Automated Annotation

Use pre-trained classifiers to automatically assign cell type labels.

Tools: CellTypist, scAnnotate Best for: Standard tissues, quick preliminary annotation, large datasets

3. Reference-Based Label Transfer

Transfer labels from annotated reference datasets to your query data.

Tools: scArches, scANVI, Azimuth, SingleR Best for: Well-characterized tissues, integration with public data

Recommended Workflow

Step 1: Quality Control First

• Remove low-quality cells before annotation
• Filter doublets (expected doublet rate: 0.8% per 1000 cells)
• Check for ambient RNA contamination
• Verify cluster quality and resolution

Step 2: Initial Marker-Based Assessment

# Scanpy example
import scanpy as sc

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

# Visualize top markers
sc.pl.rank_genes_groups(adata, n_genes=25, sharey=False)

# Plot known markers
markers = {
    'T cells': ['CD3D', 'CD3E', 'CD4', 'CD8A'],
    'B cells': ['CD19', 'MS4A1', 'CD79A'],
    'Monocytes': ['CD14', 'FCGR3A', 'LYZ'],
    'NK cells': ['NCAM1', 'NKG7', 'GNLY']
}

sc.pl.dotplot(adata, markers, groupby='leiden')

Step 3: Use Automated Tools for Validation

# CellTypist example (fast, accurate for immune cells)
import celltypist
from celltypist import models

# Download immune cell model
model = models.Model.load(model='Immune_All_Low.pkl')

# Predict cell types
predictions = celltypist.annotate(adata, model=model, majority_voting=True)
adata = predictions.to_adata()

Step 4: Reference-Based Refinement

# scArches example for label transfer
import scarches as sca

# Load pre-trained reference model
model = sca.models.SCANVI.load_query_data(
    adata=adata,  # Your query data
    reference_model="path/to/reference_model"
)

# Transfer labels
model.train(max_epochs=100)
adata.obs['transferred_labels'] = model.predict()

Best Practices

Do's:

1. Always combine multiple approaches - Use marker-based validation even with automated tools
2. Check cluster purity - Ensure clusters represent single cell types
3. Validate with multiple marker sets - Don't rely on single markers
4. Consider biological context - Tissue type, disease state, developmental stage
5. Document confidence levels - Note uncertain annotations
6. Use hierarchical annotation - Broad categories first, then subtypes

Don'ts:

1. Don't over-cluster - Too fine resolution creates artificial distinctions
2. Don't ignore batch effects - Correct before annotation
3. Don't trust automation blindly - Always validate predictions
4. Don't mix cell states with cell types - Activated vs. resting cells are states, not types
5. Don't annotate low-quality cells - Remove them first

Common Pitfalls

1. Doublet Clusters: Clusters that show markers from multiple cell types are often doublets, not novel hybrid populations.
   • How to avoid: Run doublet detection tools (Scrublet, DoubletFinder) before annotation and remove flagged cells.
2. Ambient RNA Contamination: Background markers appear across all cells, blurring cell type boundaries.
   • How to avoid: Apply SoupX or CellBender decontamination during preprocessing — don't trust raw counts on droplet data.
3. Over-interpretation of Small Clusters: Rare clusters (<25 cells) are often technical artifacts rather than biological subtypes.
   • How to avoid: Require a minimum cell count threshold and validate with an independent dataset before naming the cluster.
4. Reference Mismatch: Transferring labels from a reference built on a different tissue, species, or condition produces confidently wrong annotations.
   • How to avoid: Use tissue- and species-matched references, and check marker-gene overlap between query and reference before label transfer.
5. Confusing Cell States with Cell Types: Activated vs. resting T cells, M1 vs. M2 macrophages, and cycling vs. quiescent cells are states, not distinct types.
   • How to avoid: Annotate cell type first using stable lineage markers, then layer state annotations on top — don't mix the two axes.
6. Trusting Automated Tools Blindly: CellTypist or SingleR predictions look authoritative but can fail silently on out-of-distribution cells.
   • How to avoid: Always cross-check automated calls against marker-based dot plots, and flag low-confidence predictions for manual review.
7. Annotating Low-Quality Cells: Including cells with high mitochondrial content or low gene counts contaminates downstream signatures.
   • How to avoid: Apply QC filters (mt%, n_genes, n_counts) before clustering — don't annotate first and clean up later.

Tool Selection Guide

| Scenario | Recommended Tool | Why | |----------|------------------|-----| | Immune cells (human) | CellTypist | Pre-trained on large immune atlases | | Mouse tissues | scArches + Mouse Cell Atlas | Comprehensive mouse reference | | Novel cell types | Manual + Scanpy/Seurat | Need domain expertise | | Large datasets (>100k cells) | CellTypist | Fast, scalable | | Cross-species | Manual markers | Limited reference transfer | | Developmental data | scArches | Handles continuous states |

Key Marker Genes by Cell Type

Blood/Immune:

• T cells: CD3D, CD3E (all T cells); CD4, CD8A (subtypes)
• B cells: CD19, MS4A1 (CD20), CD79A
• Monocytes/Macrophages: CD14, CD68, LYZ
• NK cells: NCAM1 (CD56), NKG7, KLRD1
• Dendritic cells: FCER1A, CD1C

Epithelial:

• General epithelial: EPCAM, KRT18, KRT19
• Lung AT1: AGER, PDPN
• Lung AT2: SFTPC, SFTPA1
• Intestinal: VIL1, MUC2

Stromal:

• Fibroblasts: COL1A1, DCN, LUM
• Endothelial: PECAM1 (CD31), VWF, CDH5
• Smooth muscle: ACTA2, MYH11, TAGLN

Validation Checklist

• [ ] Cluster purity: >80% cells with same label per cluster
• [ ] Marker consistency: Top DE genes match expected markers
• [ ] Biological plausibility: Expected proportions for tissue type
• [ ] Cross-method agreement: Manual and automated annotations align
• [ ] Reference quality: >70% cells successfully transferred
• [ ] Doublet check: No clusters with multi-lineage markers
• [ ] Documentation: Record confidence levels and uncertain calls

References

Tools:

• Scanpy: https://scanpy.readthedocs.io/
• CellTypist: https://www.celltypist.org/
• scArches: https://scarches.readthedocs.io/
• Seurat: https://satijalab.org/seurat/

Marker Databases & Atlases:

• PanglaoDB: https://panglaodb.se/ (Database of marker genes)
• CellMarker: http://bio-bigdata.hrbmu.edu.cn/CellMarker/ (Curated cell marker database)
• Human Cell Atlas: https://www.humancellatlas.org/ (Reference datasets)
• Single Cell Best Practices: https://www.sc-best-practices.org/cellular_structure/annotation.html
• Luecken & Theis (2023): Current best practices in single-cell RNA-seq analysis. Molecular Systems Biology.

Pre-trained Models:

• CellTypist models: 30+ tissue-specific models
• Azimuth references: PBMC, lung, kidney, etc.
• scArches models: Multiple tissue references

Troubleshooting

Issue: All clusters look similar → Increase clustering resolution, check if data is normalized

Issue: Too many small clusters → Decrease resolution, merge similar clusters based on markers

Issue: Automated tool gives inconsistent results → Check input normalization, try multiple tools, fall back to manual

Issue: Can't find clear markers for cluster → May be transitional state, doublet, or low-quality cells

Issue: Reference transfer fails → Check batch correction, ensure overlapping gene sets, verify tissue match