Cell-Type Annotation

The dominant paradigm where query cells are projected onto a pre-annotated reference atlas using correlation or classification algorithms. The process compares each cell's gene expression vector against a curated database of known cell-type signatures.

• SingleR: Computes Spearman correlation between each cell and reference bulk RNA-seq datasets, iteratively refining assignments
• CellTypist: Uses logistic regression models trained on harmonized scRNA-seq references with hierarchical decision trees
• Azimuth: Leverages bridge integration to map query cells onto a reference-embedded space using anchor-based transfer learning

Reference mapping achieves high throughput but is constrained by the completeness and quality of the reference atlas.

A knowledge-driven approach that assigns cell identities based on the expression of canonical marker genes—specific genes known to characterize particular cell types. This method relies on curated gene sets from literature and databases like PanglaoDB or CellMarker.

• Manual annotation: Experts visually inspect cluster-specific differentially expressed genes against known markers
• ScType: Automated scoring system using a database of positive and negative marker gene combinations
• Garnett: Trains a hierarchical classifier from user-defined marker files, then assigns cells based on marker enrichment scores

This approach provides interpretable results but requires extensive domain expertise and may miss novel or transitional cell states.

A de novo approach where cells are grouped into clusters based on transcriptomic similarity using graph-based community detection algorithms, followed by differential expression analysis to characterize each cluster. This method enables the discovery of novel cell types not present in existing references.

• Louvain/Leiden algorithms: Partition the k-nearest neighbor graph into communities of transcriptionally similar cells
• Differential expression: Identifies cluster-specific genes using Wilcoxon rank-sum or likelihood ratio tests
• Cluster annotation: Combines automated marker detection with manual curation to assign biological labels

This approach is essential for exploratory biology but introduces subjectivity in cluster resolution and label assignment.

A structured annotation framework that maps cell identities onto established cell ontologies such as the Cell Ontology (CL) or Human Cell Atlas Ontology. This enforces standardized nomenclature and captures hierarchical relationships between cell types.

• Directed acyclic graphs (DAGs): Represent parent-child relationships (e.g., CD8+ T-cell is_a T-cell)
• CellO: Uses hierarchical classification to assign cells at multiple levels of the ontology tree
• OnClass: Leverages few-shot learning to annotate rare cell types by propagating information through the ontology structure

Ontology alignment ensures semantic interoperability across studies and enables cross-dataset meta-analysis.

A critical characteristic where annotation algorithms provide confidence scores or posterior probabilities for each cell-type assignment, distinguishing between high-confidence calls and ambiguous identities. This is essential for quality control and downstream analysis.

• Entropy-based metrics: Measure the uncertainty across all possible cell-type probability distributions
• Conformal prediction: Produces prediction sets with guaranteed coverage probabilities, flagging cells that cannot be confidently assigned
• Rejection classifiers: Implement thresholds where cells below a confidence cutoff are labeled as 'Unassigned' or 'Unknown'

Uncertainty quantification prevents false-positive annotations and identifies transitional or doublet cells that require special handling.

The process of projecting cell-type labels from a multi-modal reference (e.g., CITE-seq with transcriptome and surface proteins) onto query datasets that may contain only a single modality. This leverages the richer information in multi-omics references to improve annotation accuracy.

• Seurat v4/v5: Uses weighted nearest neighbor (WNN) analysis to integrate modalities and transfer labels via mutual nearest neighbors
• TotalVI: A variational autoencoder that learns a joint latent space for RNA and protein data, enabling label transfer to RNA-only queries
• Bridge integration: Maps single-modality queries into a multi-modal reference using dictionary learning

Cross-modality transfer enables retrospective annotation of legacy datasets using modern multi-omics atlases.