Uncertainty-aware single-cell annotation with a hierarchical reject option.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Automatic cell type annotation uses RNA-seq data to label cells.
• Gene expression features have limited resolution, causing annotation uncertainty.
• Current methods often use full rejection to handle uncertainty, but this can discard valuable information.

Purpose of the Study:

• To evaluate different rejection strategies for automatic cell type annotation.
• To compare flat versus hierarchical classifiers with various rejection approaches.
• To establish best practices for handling annotation uncertainty in RNA-seq datasets.

Main Methods:

• Evaluated three annotation approaches: full rejection, partial rejection, and no rejection.
• Compared flat and hierarchical probabilistic classifiers.
• Analyzed classifier performance based on rejection strategies and thresholding.

Main Results:

• Hierarchical classifiers outperform flat classifiers when rejection is applied.
• Partial rejection is the preferred approach, retaining significant label information.
• Without rejection, flat and hierarchical methods perform similarly if the hierarchy reflects transcriptomic relationships.
• Optimal rejection thresholds require careful examination of method-specific rejection behavior.

Conclusions:

• Hierarchical models with partial rejection offer a robust strategy for cell type annotation.
• The choice of rejection strategy and threshold significantly impacts annotation accuracy and information preservation.
• Further research into optimizing rejection mechanisms is warranted for advancing single-cell RNA-seq analysis.