scDREAMER

scDREAMER integrates heterogeneous single-cell sequencing datasets using deep generative models and adversarial training to correct batch effects while preserving biological variation and developmental trajectories.

Key Features:

• Deep Generative Models: Employs deep generative architectures to produce integrated datasets while preserving biological variation across conditions.
• Adversarial Training: Incorporates adversarial classifiers to improve robustness and accuracy of batch correction and to maintain cell type distributions when skewed among batches.
• Unsupervised and Supervised Integration (scDREAMER-Sup): Provides unsupervised integration and a supervised variant, scDREAMER-Sup, for cases with available biological labels or additional information.
• Overcoming Batch-Effects: Addresses nested batch-effects and skewed cell type distributions commonly encountered in single-cell studies.
• Scalability: Demonstrated capability to handle large-scale datasets including up to one million cells for atlas-level integration.
• Cross-Species Integration: Supports integration across species, including human and mouse datasets, for comparative analyses.
• Performance and Benchmarking: Benchmarked on six real datasets with reported superior batch-correction and conservation of biological variation and faster processing times versus other deep-learning-based integration methods.

Scientific Applications:

• Developmental Biology: Preserves developmental trajectories across time points and conditions to support studies of cellular differentiation and lineage dynamics.
• Disease Modeling: Integrates multi-batch single-cell RNA sequencing data from different tissues, time points, and conditions to enable comparative analysis of disease states.
• Cross-Species Comparative Studies: Facilitates integration of human and mouse single-cell datasets for cross-species comparisons.
• Atlas-level Integration: Enables construction of large-scale cell atlases by integrating datasets comprising up to one million cells.
• Single-cell RNA Sequencing Integration: Suited for integrating scRNA-seq datasets across batches while conserving biological signal.

Methodology:

Uses deep generative models combined with adversarial training via adversarial classifiers; supports both unsupervised integration and a supervised variant (scDREAMER-Sup); evaluated through benchmarking on six real datasets.