pscAdapt: Pre-Trained Domain Adaptation Network Based on Structural Similarity for Cell Type Annotation in Single...

Area of Science:

• Computational Biology
• Bioinformatics
• Single-cell Genomics

Background:

• Cell type annotation is vital for understanding cellular functions and biological processes.
• Automated methods often struggle with batch effects, impacting performance across different data sources.
• Batch effects arise from variations in data distribution across platforms and species.

Purpose of the Study:

• To propose pscAdapt, a pre-trained domain adaptation model for robust cell type annotation.
• To address and mitigate batch effects in single-cell data analysis.
• To enhance the accuracy and discriminability of cell type identification.

Main Methods:

• Utilized a pre-trained strategy to initialize model parameters for source domain data distribution learning.
• Employed adversarial learning to train a domain adaptation network, achieving domain-level alignment.
• Designed a structural similarity loss to align cells at the cell level, enhancing discriminability.

Main Results:

• Demonstrated the effectiveness of pscAdapt on simulated, cross-platform, and cross-species datasets.
• Showcased superior performance of pscAdapt compared to several existing popular cell type annotation methods.
• Validated the model's ability to reduce domain discrepancy and improve cell type classification.

Conclusions:

• pscAdapt effectively overcomes batch effects in cell type annotation.
• The proposed model enhances cell type discriminability through structural similarity.
• pscAdapt represents a significant advancement for accurate cell type annotation in bioinformatics.