SpaOmicsVAE: A deep learning framework for integrative analysis of spatial multi-omics data

Spatial multi-omics technologies have revolutionized our understanding of biological systems by enabling simultaneous measurement of multiple molecular features while preserving spatial information. However, the effective integration and analysis of these complex datasets remain challenging. Here we introduce SpaOmicsVAE, a novel computational framework that combines variational autoencoder architecture with a dual graph neural network to address these challenges. The framework uniquely integrates spatial and feature information through an attention-based mechanism, effectively handling data sparsity and noise while preserving crucial spatial relationships. We demonstrate SpaOmicsVAE's superior performance through comprehensive benchmarking against existing methods using both experimental and simulated datasets. Applied to diverse tissue samples including thymus, spleen, hippocampus, and brain tissues, SpaOmicsVAE revealed previously uncharacterized spatial patterns in T cell development, immune cell organization, and epigenetic regulation. Our framework provides a powerful tool for deciphering the spatial organization of complex biological systems, offering new insights into tissue architecture and function.