scbean

scbean integrates diverse single-cell datasets (scRNA-seq, scATAC-seq, and spatially resolved transcriptomics) to produce harmonized representations for downstream cellular and regulatory analyses.

Key Features:

• Domain-Adversarial Variational Autoencoder (DAVAE): Uses domain-adversarial learning and variational approximation to learn unified latent representations across samples, technologies, and modalities.
• Batch Effect Removal: Mitigates batch effects across experiments and platforms through adversarial alignment in the latent space.
• Transfer Learning: Transfers annotated cell-type labels from scRNA-seq to scATAC-seq and spatial transcriptomics to improve annotation of other modalities.
• Paired Data Integration: Integrates paired scATAC-seq and transcriptome profiles measured from the same cells to link chromatin accessibility with gene expression.
• Scalability and Efficiency: Trains with mini-batch stochastic gradient descent and supports GPU acceleration for large-scale datasets.

Scientific Applications:

• Cellular Composition Analysis: Resolves cell-type composition and distributions in heterogeneous tissues by integrating multiple single-cell datasets.
• Gene Regulation Studies: Enables investigation of gene regulatory mechanisms by linking ATAC and transcriptome data from the same cells.
• Cross-Modality Comparisons: Facilitates comparative analyses across scRNA-seq, scATAC-seq, and spatial transcriptomics modalities.

Methodology:

Implements the DAVAE model combining domain-adversarial learning with variational approximation; training uses mini-batch stochastic gradient descent and can be accelerated on GPUs.