Benchmarking cell type annotation in spatial transcriptomics: resolving cellular hierarchies, biological fidelity,...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatial transcriptomics provides gene expression data in tissue context.
• Accurate cell type annotation is essential for interpreting spatial transcriptomics data.
• Current annotation methods face challenges with fine-grained subtypes and dynamic cellular states.

Purpose of the Study:

• To systematically benchmark 20 state-of-the-art cell type annotation methods for spatial transcriptomics.
• To evaluate method performance across diverse datasets and annotation granularities.
• To provide guidance for selecting appropriate annotation strategies.

Main Methods:

• Benchmarked 20 reference-based and reference-free cell type annotation methods.
• Utilized four spatial transcriptomics datasets with expert-curated ground truth labels.
• Assessed performance using classification and structure-aware metrics.

Main Results:

• Annotation performance varied significantly based on tissue context, reference-query similarity, and annotation granularity.
• Fine-grained subtype annotation and rare cell recovery remained challenging, especially in dynamic biological processes.
• High classification accuracy did not always correlate with preserved biological organization or downstream analysis coherence.
• scANVI, Seurat, and TACCO were top performers, with context-dependent advantages.

Conclusions:

• Current cell type annotation methods for spatial transcriptomics show variable performance.
• Method selection should consider dataset characteristics, biological questions, and desired annotation resolution.
• Further development is needed to improve annotation of complex biological states and rare cell populations.