Cell Type Annotation

The traditional approach relying on curated gene sets where domain experts identify clusters by matching differentially expressed genes against known markers.

• Canonical markers: CD3E (T cells), CD14 (monocytes), EPCAM (epithelial)
• Limitation: Requires deep biological expertise and does not scale to atlas-level datasets
• Validation: Cross-reference with the CellMarker or PanglaoDB databases

Manual inspection of dot plots and violin plots confirms expression specificity before label assignment.

Projects query cells onto a well-annotated reference atlas in a shared latent space, transferring labels without manual inspection.

• Seurat v5 uses anchor-based integration to find mutual nearest neighbors
• SingleR computes Spearman correlation between each cell and reference bulk transcriptomes
• scArches enables reference mapping with architectural surgery for query-specific adaptation

Reference mapping enables standardized annotation across studies and eliminates inter-annotator variability.

Machine learning classifiers trained on labeled reference data to predict cell types in unseen datasets with probabilistic confidence scores.

• CellTypist uses logistic regression with hierarchical cell type ontologies
• scANVI extends scVI with a semi-supervised variational autoencoder
• Support Vector Machines operate directly on PCA-reduced expression space

These models output per-cell probabilities, enabling identification of ambiguous or novel states when confidence drops below thresholds.

Cell type labels exist at multiple levels of granularity, from broad lineages to fine subtypes, requiring structured ontologies for consistent annotation.

• Level 1: Major lineages (immune, stromal, epithelial)
• Level 2: Cell families (T cell, B cell, fibroblast)
• Level 3: Subtypes (CD4+ memory T cell, regulatory T cell)
• Level 4: Functional states (activated, exhausted, cycling)

The Cell Ontology (CL) provides a standardized directed acyclic graph for machine-readable cell type definitions.

Modern annotation tools provide confidence metrics that flag cells with ambiguous identity, preventing erroneous downstream biological conclusions.

• Entropy-based scores: Measure classification uncertainty across label probabilities
• Distance-based metrics: Flag cells far from reference centroids as potentially novel
• Rejection classifiers: Abstain from labeling cells below a calibrated confidence threshold

Unlabeled or low-confidence cells may represent rare populations, doublets, or transitional states requiring further investigation.

Labels can be transferred between different data modalities using shared latent space embeddings that align disparate feature spaces.

• RNA to ATAC: Bridge from gene expression to chromatin accessibility using gene activity scores
• CITE-seq integration: Use surface protein abundance as anchor features for RNA-based annotation
• Multiome references: Joint RNA+ATAC atlases enable simultaneous annotation of both modalities

Tools like Seurat WNN and totalVI learn weighted nearest-neighbor graphs across modalities for robust label transfer.

An unsupervised method that partitions cells into groups by constructing a k-nearest neighbor (KNN) graph and applying community detection. The Leiden algorithm is the current standard, optimizing the modularity of the graph partition.

• Clusters represent putative cell types before annotation
• Resolution parameter controls cluster granularity
• Directly precedes the annotation step in standard workflows

Statistical comparison of gene expression between cell groups to identify transcripts that are significantly up- or down-regulated. Methods like the Wilcoxon rank-sum test or MAST are used to find cluster-specific markers.

• Generates ranked gene lists for manual annotation
• Must account for the sparse, zero-inflated nature of scRNA-seq data
• Provides statistical confidence for marker gene assignment