scAdapt

scAdapt performs virtual adversarial domain adaptation to transfer cell-type labels between heterogeneous single-cell RNA sequencing (scRNA-seq) datasets and mitigate batch effects.

Key Features:

• Domain Adaptation: Leverages a virtual adversarial approach to align data distributions across different sources or species and mitigate batch effects.
• Enhanced Classifier Training: Trains an enhanced classifier using labeled source data and unlabeled target data with pseudo-labels to improve label-transfer robustness.
• Joint Embedding Generation: Aligns centroids from labeled source data with pseudo-labeled target data to create a joint embedding that preserves intrinsic biological structures.
• Superior Performance: Empirical evaluations demonstrate outperforming existing methods in simulated environments, cross-platform analyses, cross-species comparisons, and spatial transcriptomic studies while maintaining discriminative cluster structures and effective mixing of cells.

Scientific Applications:

• Cross-Platform Analysis: Integrates data from multiple sequencing platforms to enable combined analyses despite platform-specific batch effects.
• Cross-Species Studies: Transfers cell-type labels across species to support comparative and evolutionary biology research.
• Spatial Transcriptomics: Aligns spatial transcriptomic data with conventional scRNA-seq datasets to enhance interpretation of spatially resolved gene expression.

Methodology:

Uses a virtual adversarial domain adaptation framework, trains an enhanced classifier with labeled source data and unlabeled target data using pseudo-labels, and aligns centroids from labeled source and pseudo-labeled target to construct a joint embedding space that preserves discriminative cluster structures.