SpaBatch: Deep Learning-Based Cross-Slice Integration and 3D Spatial Domain Identification in Spatial Transcriptomics.

With the rapid accumulation of spatial transcriptomics (ST) data across diverse tissues, individuals, and technological platforms, there is an urgent need for a robust and reliable multi-slice integration framework to enable 3D spatial domain identification. However, existing methods largely focus on 2D spatial domain identification within individual slices and fail to adequately account for inter-slice spatial correlations and batch effect correction, thereby limiting the accuracy of cross-slice 3D spatial domain identification. In this study, SpaBatch is presented, a novel framework for integrating and analyzing multi-slice ST data, which effectively corrects batch effects and enables cross-slice 3D spatial domain identification. To demonstrate the power of SpaBatch, SpaBatch is applied to eight real ST datasets, including human cortical slices from different individuals, mouse brain slices generated using two different techniques, mouse embryo slices, human embryonic heart slices, HER2+ breast cancer tissues and mouse hypothalamic slices profiled using the MERFISH platforms. Comprehensive validation demonstrates that SpaBatch consistently outperforms state-of-the-art methods in 3D spatial domain identification while effectively correcting batch effects. Moreover, SpaBatch efficiently captures conserved tissue architectures and cancer-associated substructures across slices, and leverages limited annotations to predict spatial domain in unannotated sections, highlighting its potential for tissue-structure interpretation and developmental biology studies. All code and public datasets used in this study are available at: https://github.com/wenwenmin/SpaBatch.