MAT2

MAT2 aligns multiple single-cell transcriptome datasets using manifold alignment with contrastive learning to integrate scRNA-seq data and reconstruct batch-effect-free gene expression for improved cross-dataset cell type annotation and comparative cellular analysis.

Key Features:

• Manifold alignment with contrastive learning: Employs a deep neural network to perform manifold alignment using a contrastive learning strategy that leverages mutual neighbors and known cell type annotations.
• Cell triplets: Uses cell triplets derived from annotations to construct a consensus manifold that remains robust when datasets have limited common cell types.
• Batch-effect-free gene expression reconstruction: Reconstructs gene expression profiles without batch effects to support accurate cell type annotation across different datasets.

Scientific Applications:

• Benchmarking and method comparison: Demonstrated to outperform existing popular methods in benchmarking tests on real scRNA-seq datasets.
• Cross-dataset cell type annotation: Enables more accurate annotation of cell types across heterogeneous single-cell datasets.
• Comparative hematopoiesis analysis: Has been used to reveal differences in differentiation paces of hematopoietic stem cells between humans and mice.

Methodology:

Aligns single-cell transcriptomes in a manifold space using a deep neural network trained with contrastive learning, defines cell triplets from known annotations to guide alignment, and reconstructs gene expression profiles to remove batch effects.