scAdapt: virtual adversarial domain adaptation network for single cell RNA-seq data classification across platforms...

Area of Science:

• Computational Biology
• Bioinformatics
• Single-cell Genomics

Background:

• Conventional cell type identification in single-cell analysis relies on marker genes, a process that is time-consuming and prone to reproducibility issues.
• Existing supervised methods for cell type identification struggle with batch effects and biological variations across different datasets, platforms, and species.

Purpose of the Study:

• To develop a robust method, scAdapt, for transferring cell type labels between single-cell datasets, effectively addressing batch effects.
• To improve the accuracy and reliability of cell type classification in cross-dataset and cross-platform single-cell analyses.

Main Methods:

• Introduced scAdapt, a virtual adversarial domain adaptation network utilizing labeled source and unlabeled target data.
• Employed an enhanced classifier trained on both datasets and aligned labeled source centroids with pseudo-labeled target centroids.
• Generated a joint embedding space to facilitate accurate cell type transfer across datasets.

Main Results:

• scAdapt demonstrated superior performance compared to existing methods in simulated, cross-platform, cross-species, spatial transcriptomic, and COVID-19 immune datasets.
• Quantitative evaluations and visualizations confirmed scAdapt's effectiveness in cell mixing.
• The method successfully preserved the discriminative cluster structure inherent in the original datasets.

Conclusions:

• scAdapt provides a powerful solution for accurate cell type label transfer in single-cell analyses, mitigating challenges posed by batch effects and data heterogeneity.
• The developed domain adaptation network offers a significant advancement for cross-dataset cell type identification, enhancing the utility of public single-cell data.