Hierarchical progressive learning of cell identities in single-cell data

Area of Science:

• Computational Biology
• Bioinformatics
• Single-cell Genomics

Background:

• Supervised methods are crucial for cell population identification in single-cell data.
• Existing methods struggle with multi-dataset learning, varied annotation resolutions, and annotation preservation during retraining.

Purpose of the Study:

• To introduce scHPL (single-cell Hierarchical Progressive Learning), a novel method for continuous learning from single-cell data.
• To enable learning from multiple datasets with differing annotation resolutions while preserving annotations.

Main Methods:

• scHPL employs hierarchical progressive learning to build and update a classification tree.
• It leverages varying annotation resolutions across datasets for continuous learning.
• Performance is validated using both simulated and real single-cell datasets.

Main Results:

• scHPL successfully learns known cellular hierarchies from multiple datasets.
• The method demonstrates robust annotation preservation across retraining.
• Classification and tree learning performance were evaluated effectively.

Conclusions:

• scHPL offers a solution for continuous and robust learning from multi-dataset single-cell data.
• The method overcomes limitations of existing approaches in annotation handling and multi-dataset integration.
• scHPL ensures reliable downstream analysis by preserving original annotations.