scGALA advances graph link prediction-based cell alignment for comprehensive data integration and harmonization.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell technologies enable multimodal data acquisition, revealing cellular heterogeneity.
• Robust cell alignment is crucial for integrating and harmonizing single-cell data, including batch correction and multi-omics integration.
• Existing alignment methods often rely on rigid metrics, limiting accuracy across diverse cell populations.

Purpose of the Study:

• To introduce scGALA, a novel graph-based learning framework for redefining cell alignment in single-cell data integration.
• To overcome limitations of existing methods by developing a flexible and accurate cell alignment strategy.
• To enhance the performance of downstream single-cell data integration tasks.

Main Methods:

• scGALA utilizes graph attention networks and a score-driven, task-independent optimization strategy.
• It constructs enriched graphs by integrating gene expression with auxiliary data (e.g., spatial coordinates).
• Alignment is refined through self-supervised graph link prediction using a deep neural network.

Main Results:

• scGALA identifies over 25% more high-confidence cell alignments compared to existing methods.
• The framework maintains or improves accuracy in cell correspondence.
• Demonstrated versatility in enhancing various single-cell data integration tasks.

Conclusions:

• scGALA offers a significant advancement in cell alignment for single-cell data integration.
• The graph-based approach effectively captures complex cell-cell relationships.
• Improved alignment via scGALA facilitates more robust and accurate downstream analyses.