GALA: Integrating Weighted Graph Walks and Latent-Space Adversarial Training for Single-Cell Batch Alignment

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals cellular heterogeneity but suffers from batch effects.
• Technical variations across scRNA-seq datasets hinder accurate data integration and analysis.

Purpose of the Study:

• To introduce Graph-based Adversarial Latent Alignment (GALA), a novel framework for robust batch correction in scRNA-seq data.
• To align scRNA-seq datasets while preserving essential biological signals.

Main Methods:

• GALA integrates weighted graph random walks with latent space adversarial training.
• A Weighted Graph Mutual Nearest Neighbor (WGMNN) module enhances cross-batch cell pairing.
• Adversarial training in a low-dimensional latent space generates batch-agnostic representations.

Main Results:

• GALA demonstrated superior performance in batch correction across five diverse benchmark datasets, achieving high F1 scores.
• The WGMNN module improved cell pairing diversity by up to 125% and coverage by 48%.
• GALA outperformed established methods like Seurat v4, Harmony, and Scanorama, especially in complex datasets.

Conclusions:

• GALA offers a robust and computationally efficient solution for scRNA-seq data integration.
• The framework effectively eliminates technical artifacts while preserving biological variability.
• GALA shows consistent superiority and robustness across various scRNA-seq integration scenarios.