Deciphering conserved and divergent tissue architectures via soDGVAE-mediated spatial transcriptomics integration

Spatial omics has revolutionized our understanding of molecular heterogeneity within its native tissue context. However, the integrative analysis of datasets spanning replicate samples, disparate platforms, different developmental stages, and cross-species remains a challenging computational task.

Here, we present soDGVAE, a conditional graph attention variational autoencoder framework engineered for the robust harmonization and alignment of multisource spatial transcriptomics data.

By synergistically encoding gene expression and spatial topology through a graph attention variational autoencoder, soDGVAE effectively decouples technical batch effects from genuine biological variation.

Through comprehensive benchmarking, we demonstrate that the proposed model performs competitively in integrating replicate samples across platforms, large-scale multi-technology atlases, and datasets with limited gene overlap. Beyond technical integration, soDGVAE facilitates deep biological discovery: it enables the high-fidelity domain identification in the three-dimensional murine kidney architecture, and distinguishes coarse and fine-grained regions in the developmental avian heart. Notably, the cross-species study found that in the same Carnegie stage, the skin of mice matured earlier than that of humans, which elucidates the heterogeneous shifts in developmental tempo across species. Finally, we applied soDGVAE to the kidney cohort to decode the key domains.

Collectively, our results position soDGVAE as a scalable and generalizable tool, transforming discrete tissue slices into coherent, multi-dimensional biological maps.