HDMC

HDMC removes batch effects from single-cell RNA sequencing (scRNA-seq) data by combining hierarchical distribution matching with contrastive learning to align batches while preserving biological variability.

Key Features:

• Hierarchical distribution matching: Uses a deep autoencoder-based hierarchical framework with adversarial training to align global distributions across batches.
• Local distribution alignment (MMD): Employs a maximum mean discrepancy (MMD)-based loss with within-batch clustering to match local distributions.
• Contrastive learning: Integrates contrastive learning to align similar cluster pairs and separate dissimilar ones, reducing false-positive cluster merges.
• Overcorrection prevention: Preserves genuine biological variability while removing technical batch effects.

Scientific Applications:

• Batch effect correction in scRNA-seq: Reduces distribution differences between batches and aligns samples to recover cell type clusters.
• Integration of simulated and real datasets: Demonstrated effectiveness on both simulated and real scRNA-seq datasets, outperforming state-of-the-art methods.
• Preservation of biological signal: Maintains cell-type-specific biological variability during dataset integration, mitigating overcorrection.

Methodology:

Deep autoencoder-based hierarchical framework; adversarial training for global distribution alignment; maximum mean discrepancy (MMD)-based loss with within-batch clustering for local alignment; contrastive learning to align similar cluster pairs and separate dissimilar ones, minimizing false positives and preventing overcorrection.