Abstract
Single-cell technologies have transformed our understanding of cellular heterogeneity through multimodal data acquisition. However, robust cell alignment remains a major challenge for data integration and harmonization, including batch correction, label transfer, and multi-omics integration. Many existing methods constrain alignment based on rigid feature-wise distance metrics, limiting their ability to capture accurate cell correspondence across diverse cell populations and conditions. We introduce scGALA, a graph-based learning framework that redefines cell alignment by combining graph attention networks with a score-driven, task-independent optimization strategy. scGALA constructs enriched graphs of cell-cell relationships by integrating gene expression profiles with auxiliary information, such as spatial coordinates, and iteratively refines alignment via self-supervised graph link prediction, where a deep neural network is trained to identify and reinforce high-confidence correspondences across datasets. In extensive benchmarks, scGALA identifies over 25 percent more high-confidence alignments without compromising accuracy. By improving the core step of cell alignment, scGALA serves as a versatile enhancer for a wide range of single-cell data integration tasks.