Cell Type Annotation

The foundational approach where domain experts assign cell identities by examining the expression of canonical marker genes within each cluster.

• Workflow: Generate cluster-specific differentially expressed genes, then cross-reference against known markers (e.g., CD3E for T cells, MS4A1 for B cells)
• Tools: Seurat's FindAllMarkers(), Scanpy's rank_genes_groups()
• Limitation: Requires deep biological expertise and does not scale to hundreds of clusters

A supervised approach that projects cell-type labels from a well-curated reference atlas onto a query dataset by identifying transcriptional nearest neighbors.

• Label Transfer: Algorithms like Seurat v3/v4 and SingleR compute similarity scores between query cells and reference profiles
• Key Requirement: The reference must cover the expected biological diversity; mismatched tissues produce spurious annotations
• Confidence Scoring: Most tools return a prediction score per cell, enabling filtering of ambiguous assignments

Annotation strategies that leverage structured biological ontologies to assign identities at multiple resolutions, from broad lineages to fine subtypes.

• Cell Ontology (CL): A controlled vocabulary of cell types with directed acyclic graph relationships (e.g., CL:0000084 → T cell)
• Methods: Tools like CellO and OnClass use hierarchical classifiers that first distinguish major lineages before refining to terminal cell states
• Advantage: Resolves ambiguous cases by backing off to higher-level terms when fine-grained confidence is low

Neural network models trained on large-scale reference corpora to perform cell-type classification directly from raw or normalized expression vectors.

• Architectures: Multi-layer perceptrons, attention-based models, and graph neural networks that learn cell-type-discriminative features
• scANVI: A semi-supervised variational autoencoder that jointly models batch effects and cell-type labels in a probabilistic latent space
• Benefit: Captures non-linear expression patterns that linear correlation methods miss, improving accuracy on rare and transitional populations

A meta-strategy that combines predictions from multiple independent annotation methods to produce a high-confidence consensus label for each cell.

• Voting Schemes: Majority vote, weighted confidence aggregation, or Bayesian consensus across tools like SingleR, CellTypist, and scmap
• Discordance Resolution: Cells with conflicting annotations are flagged for manual review or assigned to an 'unresolved' category
• Implementation: The scConsensus framework systematically merges outputs, reducing the error rate of any single method

Emerging approaches using large pretrained single-cell foundation models that annotate cells without explicit reference mapping or retraining.

• Geneformer: A context-aware transformer pretrained on ~30M single-cell transcriptomes; cell-type identity emerges from its embedding space via similarity search
• scGPT: A generative pretrained transformer that can perform cell-type annotation as a downstream task through fine-tuning or in-context learning
• Mechanism: These models encode a universal cell representation where biologically similar cells naturally cluster, enabling annotation by embedding proximity

A supervised machine learning approach that projects cell-type annotations from a well-characterized reference atlas onto a new query dataset by identifying transcriptional similarities across studies. This method bypasses manual marker gene curation, using canonical correlation analysis or mutual nearest neighbors to align query cells with reference labels. It is essential for harmonizing annotations across experiments and scaling atlas-level analysis.

Curated repositories of cell-type-specific genes that serve as the ground truth for annotation. These databases catalog canonical markers like CD3D for T cells or MS4A1 for B cells. Automated pipelines score clusters against these panels using statistical tests such as the Wilcoxon rank-sum test or logistic regression classifiers, assigning identities based on the highest combined enrichment score.

A graph-based community detection algorithm that partitions single-cell neighborhoods into biologically meaningful clusters. It improves upon the Louvain method with a guaranteed well-connected community structure and faster convergence. The resolution parameter directly controls the granularity of clusters, making it a critical preprocessing step before annotation, as the resulting partitions define the units that receive biological labels.

Large-scale pretrained transformer models like Geneformer and scGPT that learn universal cell representations from massive single-cell corpora. These context-aware models can perform zero-shot cell type annotation by fine-tuning on small labeled datasets or by querying against reference embeddings. They capture nuanced gene-gene interaction networks, enabling annotation of rare and transitional cell states that traditional marker-based methods often miss.

The computational projection of new single-cell profiles onto an established reference atlas to rapidly annotate cell types without full dataset reintegration. Tools like Azimuth and scArches use reference-based latent space alignment, enabling real-time annotation of streaming data. This approach identifies novel cell states as those that fail to map confidently, flagging them for expert review.

Single-Cell rEgulatory Network Inference and Clustering, a method that identifies active transcription factors and their target regulons by combining co-expression analysis with cis-regulatory motif enrichment. Unlike surface-marker annotation, SCENIC assigns cell types based on the underlying gene regulatory network activity, providing a mechanistic validation of computationally derived identities and revealing master regulators of cell state.