single-cell-annotation-guide

What it does

The skill provides a structured, three-tier decision framework for annotating cell types in single-cell RNA-seq data: Tier 1 manual marker-based annotation, Tier 2 automated CellTypist, and Tier 3 reference-based popV ensemble label transfer. It explains when to use each tier, how to combine methods, and how to validate annotations with marker visualization and confidence measures. It also lists best practices and common pitfalls, along with concrete steps for QC and workflow decisions.

How it works

The guide outlines concrete decision logic and procedures, including:

• Planning annotation strategy by dataset size and tissue context using a decision tree and scenarios.
• Recommended practices such as removing doublets with Scrublet or DoubletFinder before annotation.
• Visual validation of canonical markers via dotplots, violin plots, or UMAP feature plots after automated annotation.
• Using multiple methods (e.g., CellTypist and popV) and interpreting their agreement via a method-agreement score.
• Selecting tissue-matched reference atlases (e.g., HLCA) for label transfer and applying batch correction before transfer (Harmony, scVI, BBKNN).
• Adjusting annotation granularity with clustering resolution and iterating between clustering and annotation.
• Flagging low-confidence cells based on per-cell metrics (CellTypist confidence, popV score) for manual review.
• Best practices and pitfalls with explicit actions to avoid misannotation.

When to use it

Use when planning or troubleshooting annotation in scRNA-seq experiments. The framework provides explicit scenarios (small datasets, large immune datasets, cross-atlas integration, fetal/developmental contexts, etc.) and prescribes the tier to apply in each case.

What it can touch

It references tools and scripts explicitly, including: Scrublet, DoubletFinder, sc.pl.dotplot, sc.pl.violin, sc.pl.umap, CellTypist, popV, Harmony, BBKNN, scVI, Leiden clustering, and visualization and QC steps. It specifies actions like applying doublet detection before clustering, visualizing canonical markers after annotation, computing and comparing method agreement scores, and performing batch correction prior to label transfer.

Caveats

License is CC-BY-4.0. The guide emphasizes that automated methods can misannotate and recommends validation with canonical markers; it notes that label transfer quality depends on tissue-reference match and developmental stage. It advises using multiple methods and reporting confidence thresholds; it does not guarantee correct annotations and stresses careful interpretation and documentation of methods.