Context-Aware Self-Training Framework for Cell Type Annotation Using Marker Genes

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell annotation is crucial for analyzing single-cell data.
• Self-training strategies improve annotation accuracy but require tailored frameworks for single-cell applications.

Purpose of the Study:

• To develop a context-aware self-training model (CSSTA) for improved single-cell data annotation.
• To enhance the compatibility of marker genes and optimize self-training strategies for single-cell data.

Main Methods:

• CSSTA incorporates contextual information of marker genes to generate high-quality pseudo-labels.
• It employs distinct high- and low-confidence pseudo-label recognition and supervision strategies.
• The model integrates cell-cell association insights from single-cell foundation models.

Main Results:

• Contextual marker gene information significantly improves cell-cell type association recognition.
• CSSTA outperforms existing state-of-the-art methods in benchmark experiments.
• The model demonstrates potential for hierarchical cellular annotation.

Conclusions:

• CSSTA offers a more effective self-training framework for single-cell annotation.
• The integration of marker gene context and advanced pseudo-labeling strategies enhances model performance.
• CSSTA shows promise for complex hierarchical annotation tasks in single-cell biology.