Cross-dataset annotation harmonization for cell-type hierarchy construction.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell transcriptomic datasets exhibit diverse cell-type annotation schemes and resolutions.
• This heterogeneity complicates the creation of unified cell-type hierarchies and large-scale dataset integration.
• Existing computational methods aim to harmonize annotations and build data-driven cell-type hierarchies.

Purpose of the Study:

• To benchmark existing state-of-the-art methods for cell-type annotation harmonization and hierarchy construction.
• To develop a novel computational tool for improved cell-type harmonization and hierarchy construction across datasets.

Main Methods:

• Benchmarking of scHPL, treeArches, and CellHint across simulated and real-world datasets.
• Development of three metrics: Annotation Harmonization F1-score (AH-F1), Tree Edit Distance Similarity (TEDS), and Parent-Children Branches Similarity (PCBS).
• Implementation of OTHarmonizer using partial optimal transport (OT) for cell-type harmonization and hierarchy construction.

Main Results:

• Existing methods performed well in simulated scenarios but showed limitations in complex real-world data.
• OTHarmonizer demonstrated superior accuracy in capturing equivalent and hierarchical cell-type relationships.
• The developed metrics effectively assessed the quality of constructed cell-type hierarchies.

Conclusions:

• OTHarmonizer provides a more effective approach for harmonizing cell-type annotations and constructing hierarchies across diverse single-cell datasets.
• The study highlights the need for improved computational tools to handle the complexity of real-world single-cell data.
• The benchmark and OTHarmonizer tool are publicly available to facilitate further research.