SpaVCCA Identifies Spatial Transcriptomics Domains Across Slices by Coupling Variational Autoencoder with Canonical Correlation Analysis

Spatial transcriptomics enables the characterization of gene expression within intact tissue architecture, but identifying domains across batches, platforms, or conditions remains challenging due to technical variability and batch effects. Existing methods often fail to simultaneously achieve effective batch correction and preservation of spatial structure. Here, we propose SpaVCCA, a unified framework for spatial transcriptomics integration that combines a graph convolutional Variational Autoencoder (VAE) with a Canonical Correlation Analysis (CCA)-based alignment loss and a graph contrastive objective. The graph convolutional encoder captures local spatial dependencies, while the CCA loss aligns latent representations across batches. The contrastive loss further preserves local neighborhood structure during integration. We evaluate SpaVCCA on diverse data sets, including different platforms (Visium, Stereoseq, MERFISH) with multiple slices, cross-technology integration, and 3D spatial data. Compared to existing state-of-the-art (SOTA) methods, SpaVCCA delivers substantial performance advancements. Overall, SpaVCCA provides a scalable and effective solution for integrating spatial transcriptomics data, characterizing spatially organized cellular states and supporting applications in drug development and disease research.