Single-Cell Annotation: A Workflow for Cell Type Identification

Cell type annotation is the process of assigning biological identities to clusters of cells in single-cell RNA sequencing (scRNA-seq) data. This workflow provides a systematic approach using marker genes, reference-based mapping, and manual curation, with a checklist for validation. Researchers, analysts, and life-science professionals can apply this framework to move from raw clustering output to biologically meaningful cell labels that support downstream analysis.

At a Glance

Cell type annotation bridges the gap between computational clustering and biological interpretation. The table below summarizes the primary annotation approaches, their inputs, and practical considerations for selection.

Understanding the Annotation Problem

Single-cell RNA sequencing enables characterization of cellular heterogeneity at unprecedented resolution, allowing identification of rare cell types and exploration of cell-cell communication [5]. The technology has broad applications in basic and clinical research, but the reliability of cell annotation remains a central challenge [5]. Annotation errors propagate through downstream analyses, affecting interpretation of cell subpopulations and differential expression results [5].

The annotation problem has several layers. First, clustering algorithms partition cells into groups based on transcriptional similarity, but these groups do not always correspond to distinct biological cell types. Second, even when clusters are biologically meaningful, assigning the correct label requires knowledge of cell-type-specific markers. Third, technical artifacts such as doublets can create clusters that appear to be real cells but are not [6].

A practical workflow addresses each layer systematically. The workflow begins with quality control, proceeds through clustering and marker identification, applies annotation methods, and concludes with validation and refinement.

Core Principles of Cell Type Annotation

Marker Genes Define Cell Identity

Marker genes are genes with expression restricted to specific cell types. The identification of cell types relies on the premise that transcriptional programs are sufficiently distinct to separate cell populations. Traditional annotation first clusters cells using unsupervised learning, then labels clusters using aggregated expression profiles and marker gene information [19]. This procedure depends heavily on clustering quality, and false detection of cluster features can lead to incorrect annotations [19].

Marker-based annotation requires a curated marker map. One resource, the ACT web server, was constructed by manually curating over 26,000 cell marker entries from approximately 7,000 publications [20]. This hierarchical marker map supports weighted gene set enrichment analysis to integrate canonical markers with ordered differentially expressed genes [20]. Such resources reduce the burden of manual literature review and provide a starting point for annotation.

Reference Data Provides Context

Reference-based annotation maps query cells to an annotated reference dataset. This approach is powerful because it leverages existing expert annotations and enables consistent labeling across studies. However, most reference datasets are constructed from single-cell RNA-sequencing data and cannot directly annotate datasets that do not measure gene expression [10].

Recent advances address this limitation. Bridge integration uses a multiomic dataset as a molecular bridge to integrate single-cell datasets across modalities [10]. Each cell in the multiomic dataset constitutes an element in a dictionary used to reconstruct unimodal datasets and transform them into a shared space [10]. This approach accurately integrates transcriptomic data with independent measurements of chromatin accessibility, histone modifications, DNA methylation, and protein levels [10].

Annotation Quality Depends on Data Quality

Annotation accuracy is limited by the quality of the input data. Doublets form when two cells are encapsulated into one reaction volume, and they appear as real cells in downstream analysis [6]. Computational doublet detection methods vary in accuracy and efficiency, with DoubletFinder showing the best detection accuracy and cxds showing the highest computational efficiency in a benchmark study [6]. Removing doublets before clustering reduces the risk of annotating artifact clusters.

Sparsity and low expression also challenge annotation [5]. Dropout events create zeros in the expression matrix, making it difficult to distinguish true absence of expression from technical failure. Methods that model count distributions and account for uncertainty caused by biological and measurement noise improve annotation reliability [7].

Practical Workflow for Cell Type Annotation

Step 1: Quality Control and Preprocessing

Quality control removes low-quality cells and technical artifacts before clustering. Filter cells based on the number of detected genes, total counts, and mitochondrial read fraction. The specific thresholds depend on tissue type and sequencing platform, so examine distributions instead of applying fixed cutoffs.

Doublet detection should occur early in the workflow. Computational methods identify cells that likely represent two captured cells, and removing these cells prevents artifact clusters from being annotated as real cell types [6]. Record the number of cells removed at each quality control step to document the analysis.

Step 2: Normalization and Feature Selection

Normalization adjusts for differences in sequencing depth across cells. Several approaches exist, and the choice affects downstream clustering and annotation. After normalization, select highly variable genes for clustering. This step reduces noise and computational burden while preserving biological signal.

Step 3: Dimensionality Reduction and Clustering

Principal component analysis reduces the dimensionality of the expression matrix. The number of principal components retained influences clustering resolution. Graph-based clustering methods such as Leiden and Infomap partition cells into groups based on shared expression patterns [17]. A comparative analysis found that clusters identified by multiresolution Infomap and Leiden show closer alignment, with Infomap offering effective characterization of cellular landscapes [17].

Clustering resolution determines the granularity of cell groups. Low resolution produces broad cell types, while high resolution separates subtypes and states. Run clustering at multiple resolutions and examine cluster stability before annotation.

Step 4: Marker Gene Identification

For each cluster, identify differentially expressed genes compared to all other clusters. These genes serve as candidate markers for annotation. Ordered lists of differentially expressed genes provide input for marker-based annotation methods [20].

Cluster purity affects marker identification. If clusters contain multiple cell types, the aggregated expression profile may not represent any single cell type accurately [19]. Examine cluster-level expression of known markers to assess purity before assigning labels.

Step 5: Annotation Method Selection

Choose an annotation method based on data characteristics, available references, and analysis goals. The decision tree below guides method selection.

Decision Tree for Annotation Tool Selection

1. Is a well-annotated reference dataset available for the same tissue and condition?

   • Yes: Use reference-based mapping for initial annotation, then validate with markers.
   • No: Proceed to marker-based annotation.

2. Are curated marker lists available for the tissue of interest?

   • Yes: Use marker gene scoring methods such as ACT or ACAM.
   • No: Use unsupervised clustering followed by manual curation with literature review.

3. Is the dataset large or multimodal?

   • Large datasets: Consider scalable methods such as dictionary learning with sketching [10].
   • Multimodal data: Use bridge integration to map across modalities [10].

4. Is expert time limited?

   • Yes: Consider automated methods, including large language model assistance [8].
   • No: Manual curation provides the highest control over annotation quality.

Step 6: Manual Curation and Validation

Automated annotation provides initial labels, but manual curation remains essential for accuracy. Review cluster markers in the context of tissue biology, check for known marker expression patterns, and refine labels where automated methods disagree with biological expectation.

Expert consensus validation strengthens annotation confidence. In a large integrated atlas of human atherosclerotic plaques, robust cell type annotations were validated by expert consensus and surface protein measurements [9]. This multi-level validation approach identifies annotation errors that single-method approaches miss.

Marker-Based Annotation Methods

Cluster-Level Marker Scoring

Cluster-level marker scoring aggregates expression within clusters and compares against curated marker lists. The ACAM method delineates a framework for automatic annotation through representative cluster identification, representative cluster annotation using marker genes, and classification of remaining cells [19]. This approach improves on traditional methods by addressing cluster purity issues [19].

The ACT web server provides a user-friendly interface for marker-based annotation [20]. Users input a list of upregulated genes, and the server returns interactive hierarchy maps with statistical information [20]. The hierarchical marker map supports annotation of less-studied tissues through a pan-tissue marker map [20].

Graph-Based and Network Methods

Graph-based methods use cell-cell networks derived from gene expression data for clustering and annotation [17]. Network clustering algorithms such as Infomap and Stochastic Block Models provide alternatives to standard approaches [17]. These methods offer insights for precise characterization of cellular landscapes in neurodegeneration and immunology [17].

Graph convolutional network methods extend this approach to annotation. A graph-based single cell type annotation method, gCAnno, uses graph structures to propagate labels from known to unknown cells [21]. More recent graph neural network approaches, including GCN-based methods, apply similar principles with deep learning architectures [23].

Deep Learning Approaches

Deep generative models provide probabilistic frameworks for annotation. The scVI method offers a fully probabilistic approach for joint representation and analysis of scRNA-seq data while accounting for uncertainty from biological and measurement noise [7]. The semi-supervised variant scANVI leverages existing cell state annotations to annotate new datasets [7]. Both methods integrate multiple datasets with a single generative model usable for downstream tasks [7].

Large language models have emerged as annotation tools. GPT-4 can accurately annotate cell types using marker gene information, generating annotations with strong concordance to manual annotations across hundreds of tissue and cell types [8]. The GPTCelltype R package implements this capability for automated annotation [8]. This approach reduces the effort and expertise required for cell type annotation [8].

Transformer-based architectures address scalability. scSwinTNet applies shifted window attention for cell type annotation in large-scale single-cell RNA-seq data [22]. These methods handle the computational demands of datasets with millions of cells.

Reference-Based Mapping Approaches

Integration and Label Transfer

Reference-based mapping integrates query cells with annotated reference cells in a shared space, then transfers labels from reference to query. This approach ensures consistency with existing annotations and enables comparison across datasets.

The scVI and scANVI methods integrate multiple datasets with a single generative model [7]. This integration accounts for batch effects and technical differences while preserving biological signal [7]. The probabilistic framework provides uncertainty estimates for annotations [7].

Dictionary Learning and Bridge Integration

Dictionary learning represents query cells as combinations of reference cells. Each cell in a multiomic dataset constitutes an element in a dictionary used to reconstruct unimodal datasets [10]. This approach transforms datasets into a shared space, enabling annotation across modalities [10].

Bridge integration addresses the limitation that most references are transcriptomic. By using a multiomic dataset as a molecular bridge, this method integrates transcriptomic data with chromatin accessibility, histone modifications, DNA methylation, and protein measurements [10]. Combined with sketching techniques, dictionary learning harmonized 8.6 million human immune cell profiles from sequencing and mass cytometry experiments [10].

Reference Atlases

Large integrated atlases provide high-quality references for annotation. An integrated single-cell atlas of human atherosclerotic plaques covers roughly 250,000 high-quality annotated cells [9]. The atlas enables accurate automatic annotation of new datasets, guides experimental design by informing sample size and detection power, and supports deconvolution of bulk RNA-seq data [9].

Reference atlases require consistent annotations to be useful. Existing datasets often lack comprehensive coverage and consistent annotations, limiting the utility of downstream analyses [9]. When constructing or using a reference atlas, verify that annotations were validated through expert consensus and orthogonal measurements [9].

Spatial Context and Multimodal Integration

Spatial Transcriptomics Integration

Spatial transcriptomics adds tissue context to single-cell annotations. Microarray-based spatial transcriptomics reveals spatial patterns of gene expression using arrays of spots, each capturing transcriptomes of multiple adjacent cells [11]. Combining spatial data with scRNA-seq from the same sample enables annotation of the precise cellular composition of distinct tissue regions [11].

Multimodal intersection analysis maps scRNA-seq-defined subpopulations to spatial locations [11]. Applied to pancreatic tumors, this approach found that subpopulations of ductal cells, macrophages, dendritic cells, and cancer cells have spatially restricted enrichments and distinct coenrichments with other cell types [11]. Spatial validation strengthens annotation confidence by confirming that annotated cell types localize to expected tissue regions.

Epigenomic Integration

Chromatin state annotation provides complementary evidence for cell identity. ChIP-seq analysis enables genome-wide analysis of histone modifications, enhancer analysis, and chromatin state annotation [12]. These approaches reveal how the epigenomic landscape contributes to cell identity, development, lineage specification, and disease [12].

Single-cell ChIP-seq methods elucidate cellular diversity within complex tissues and cancers [12]. Integrating epigenomic data with transcriptomic annotations can resolve cell states that are transcriptionally similar but epigenomically distinct.

Records and Measurements for Annotation

Documentation Requirements

Maintain records of each annotation decision to support reproducibility and interpretation. The following records should be preserved for every dataset:

Quality Metrics

Track quantitative metrics to assess annotation quality. Cluster separation metrics indicate how distinct clusters are in expression space. Marker gene enrichment scores show how strongly cluster expression matches expected marker patterns. Reference concordance measures agreement between query annotations and reference labels.

For reference-based methods, examine the confidence scores assigned to each cell. Low-confidence annotations warrant manual review. For marker-based methods, examine the statistical significance of marker enrichment in each cluster.

Common Failure Patterns in Cell Type Annotation

Cluster Purity Failures

Clusters containing multiple cell types produce ambiguous marker profiles. This failure occurs when clustering resolution is too low or when rare cell types are merged with abundant neighbors. Examine cluster-level expression of known markers to detect mixed clusters. Increase clustering resolution or apply subclustering to resolve mixed populations.

Reference Mismatch

Reference datasets may not cover all cell states present in the query data. Disease conditions, developmental stages, and species differences introduce cell states absent from references. When reference-based annotation assigns low-confidence labels or fails to classify cells, supplement with marker-based annotation and manual curation.

Doublet Artifacts

Doublets create clusters that appear to be real cell types with mixed expression profiles [6]. These artifact clusters often express markers of multiple cell types simultaneously. Run doublet detection before clustering and examine suspicious clusters for co-expression of mutually exclusive markers.

Batch Effects

Technical differences between datasets create clusters based on batch instead of biology. Integration methods such as scVI and scANVI account for batch effects while preserving biological signal [7]. Without proper integration, annotation may reflect technical artifacts instead of cell identity.

Overannotation

Assigning labels at finer resolution than supported by data creates spurious cell types. This failure occurs when clustering resolution is too high or when markers lack specificity. Validate subtype annotations with orthogonal evidence such as surface protein measurements or spatial localization [9].

Limitations and Interpretation Boundaries

Annotation Is Inferential

Cell type labels are inferences based on transcriptional similarity to known cell types. Transcriptional programs overlap between related cell types, and novel cell states may not match existing annotations. Interpret annotations as hypotheses to be tested instead of definitive identities.

Marker Specificity Is Context-Dependent

Marker genes show tissue and condition-specific expression. A marker that distinguishes cell types in one tissue may not apply in another. Validate marker choices against tissue-specific literature and reference atlases.

Reference Bias

Reference-based annotation inherits biases from the reference dataset. If the reference lacks certain cell types or states, query cells in those states will be misannotated or left unlabeled. Assess reference coverage before relying on label transfer.

Species Differences

Annotation tools and markers developed for human or mouse data may not transfer directly to other species. Porcine single-cell studies face challenges including incomplete genome annotation and limitations in robust cell-type annotation [15]. When annotating non-model species, validate markers against species-specific expression patterns.

Safety and Regulatory Context

Data Sharing and Privacy

Single-cell datasets may contain sensitive information, particularly when derived from human subjects. The NIH Genomic Data Sharing Policy governs the sharing of genomic data generated from NIH-funded research [3]. Review applicable policies before depositing or sharing single-cell data.

Data Repositories

Public repositories support data sharing and reproducibility. The NCBI provides data resources for genomic and biomedical research [2]. The EMBL-EBI Training program offers educational resources for bioinformatics analysis [1]. Depositing annotated datasets in public repositories enables validation and reuse by the research community.

FAIR Data Principles

The FAIR Guiding Principles describe standards for findability, accessibility, interoperability, and reusability of data [4]. Applying these principles to single-cell annotations includes using standard cell type ontologies, documenting annotation methods, and providing machine-readable annotation files.

Professional Escalation Criteria

When to Seek Expert Review

Escalate to a domain expert or bioinformatics specialist when annotation results have high stakes or when automated methods produce unexpected results. Specific situations warranting escalation include:

1. Annotations that will guide clinical decisions or therapeutic targeting
2. Novel cell types or states not described in existing references
3. Persistent disagreement between annotation methods
4. Low-confidence annotations in clusters with biological importance
5. Cross-species annotation where marker transfer is uncertain

When to Revisit the Workflow

Return to earlier workflow steps when validation reveals systematic annotation problems. If multiple clusters fail marker validation, revisit clustering parameters. If reference-based annotation produces widespread low-confidence labels, assess reference coverage and consider alternative references. If doublet detection was skipped, run it before re-annotating.

Validation Checklist

Use this checklist to validate annotations before proceeding to downstream analysis:

1. Quality control removed low-quality cells and doublets
2. Clustering was run at multiple resolutions with stable clusters identified
3. Marker genes for each cluster match expected expression patterns
4. Automated annotations were reviewed by a domain expert
5. Reference-based annotations include confidence scores
6. Spatial or protein-level validation confirms key annotations
7. Annotation decisions are documented with marker sources
8. Cell type proportions match expected tissue composition
9. Rare cell types are supported by sufficient cell numbers
10. Batch effects were assessed and addressed

Frequently Asked Questions

What is the difference between marker-based and reference-based annotation?

Marker-based annotation uses curated lists of genes known to be expressed in specific cell types to score clusters or cells. Reference-based annotation maps query cells to an annotated reference dataset and transfers labels from the reference. Marker-based methods work without reference data but depend on marker quality and cluster purity [19]. Reference-based methods leverage existing expert annotations and provide consistent labeling across datasets [7].

How do I choose between different annotation tools?

Consider reference availability, dataset size, modality, and available expertise. If a well-annotated reference exists for your tissue and condition, reference-based mapping provides consistent labels. If no reference exists, marker-based methods such as ACT or ACAM provide interpretable results [19][20]. For large datasets, scalable methods such as dictionary learning with sketching handle millions of cells [10]. For multimodal data, bridge integration enables cross-modality annotation [10].

What causes annotation errors in single-cell data?

Annotation errors arise from cluster impurity, doublet artifacts, batch effects, and reference mismatch. Clusters containing multiple cell types produce ambiguous marker profiles [19]. Doublets appear as real cells with mixed expression [6]. Batch effects create clusters based on technical variation instead of biology. References that lack relevant cell states produce misannotations for those states.

How many marker genes do I need for reliable annotation?

The number of markers depends on the specificity of the markers and the similarity between cell types. Using multiple markers per cell type reduces the risk of misannotation from noisy expression. Resources such as ACT integrate prevalence of canonical markers with ordered differentially expressed genes to improve reliability [20]. Validate marker choices against tissue-specific literature.

Can large language models replace manual annotation?

Large language models such as GPT-4 can accurately annotate cell types using marker gene information, with strong concordance to manual annotations across hundreds of tissue and cell types [8]. These models reduce the effort and expertise required for annotation [8]. However, manual curation remains valuable for context-specific biology and for validating automated results.

How do I annotate cell types in a non-model species?

Non-model species present challenges including incomplete genome annotation and limited marker resources [15]. Start with marker-based annotation using orthologous markers from model species, then validate with species-specific expression patterns. Reference atlases for the species, when available, provide the most reliable annotation basis.

What should I do when clusters do not match any known cell type?

First verify cluster quality by examining marker expression and checking for doublet artifacts. If the cluster appears real, consider whether it represents a novel cell state or subtype. Search for markers in the literature and compare with related tissues. If the cluster remains unidentifiable, label it as unknown and document the evidence for its distinct identity.

How do I validate annotations without spatial or protein data?

Without orthogonal measurements, validation relies on marker expression patterns, cluster stability, and expert review. Examine whether markers for each annotated cell type are expressed in the expected cells. Assess whether cell type proportions match expected tissue composition. Compare annotations with published datasets from the same tissue. Expert consensus provides additional confidence [9].

Related Bioinformatics Guides

• Single-Cell RNA-seq Clustering and Cell-Type Annotation Pipelines
• Foundation Models for Single-Cell Biology
• Single-Cell ATAC-Seq Bioinformatics
• Single-Cell RNA Sequencing: From Bulk to Resolution
• Master Guide: Single-Cell RNA Sequencing Bioinformatics Workflows

References and Further Reading

• EMBL-EBI Training. European Bioinformatics Institute.
• NCBI Data Resources. National Center for Biotechnology Information.
• Genomic Data Sharing Policy. National Institutes of Health.
• The FAIR Guiding Principles. Scientific Data.
• A Review of Single-Cell RNA-Seq Annotation, Integration, and Cell-Cell Communication.. Cells, 2023.
• Benchmarking Computational Doublet-Detection Methods for Single-Cell RNA Sequencing Data.. Cell systems, 2021.
• Probabilistic harmonization and annotation of single-cell transcriptomics data with deep generative models.. Molecular systems biology, 2021.
• Assessing GPT-4 for cell type annotation in single-cell RNA-seq analysis.. Nature methods, 2024.
• Integrated single-cell atlas of human atherosclerotic plaques.. Nature communications, 2025.
• Dictionary learning for integrative, multimodal and scalable single-cell analysis.. Nature biotechnology, 2024.
• Integrating microarray-based spatial transcriptomics and single-cell RNA-seq reveals tissue architecture in pancreatic ductal adenocarcinomas.. Nature biotechnology, 2020.
• Methods for ChIP-seq analysis: A practical workflow and advanced applications.. Methods (San Diego, Calif.), 2021.
• Identification of Potential Biomarkers for Rheumatoid Arthritis Based on Integrated Bioinformatics and Single-Cell RNA-Seq.. 2026.
• Transcriptomic characterization of key psoriasis-associated genes based on single-cell RNA-seq and machine learning.. 2026.
• Single-Cell RNA Sequencing in Porcine Biology and Production.. 2026.
• Identification and experimental validation of CD74, PGLYRP1, and TXN as potential biomarkers in rheumatoid arthritis: an integrative bulk and ScRNA-seq study.. 2026.
• Network clustering algorithms and preprocessing pipelines for robust cell type identification in single-cell RNA sequencing data.. 2026.
• Context-Aware Self-Training Framework for Cell Type Annotation Using Marker Genes.. IEEE transactions on computational biology and bioinformatics, 2026.
• Automatic Cell Type Annotation Using Marker Genes for Single-Cell RNA Sequencing Data. Biomolecules, 2022.
• Annotation of cell types (ACT): a convenient web server for cell type annotation. Genome Medicine, 2023.
• gCAnno: a graph-based single cell type annotation method. BMC Genomics, 2020.
• scSwinTNet: A Cell Type Annotation Method for Large-Scale Single-Cell RNA-Seq Data Based on Shifted Window Attention. IEEE Journal of Biomedical and Health Informatics, 2025.
• A Novel GCN-Based Cell Annotation Method for Single-Cell RNA Sequencing Data. Proceedings 2024 17th International Congress on Image and Signal Processing Biomedical Engineering and Informatics Cisp Bmei 2024, 2024.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.