STUltra: scalable and accurate integration for subcellular-level spatial omics data

Abstract

High-resolution spatial transcriptomics (ST) is generating massive datasets involving multiple tissue sections, continuous developmental stages, or various disease conditions. Yet, their scale, biological heterogeneity, and substantial batch effects hinder data integration for further analysis, particularly the temporal interpretation of disease-associated spatial domains within tissue organization landscapes. Here, we present STUltra, a scalable hypergraph framework that integrates multi-sample ST data for precise domain detection and genome-wide association study (GWAS)-based disease trait mapping. From ST datasets, STUltra first constructs interval-sampled integrative hypergraphs, in which hyperedges capture tissue neighborhoods within the slices as well as shared biological contexts across the slices. It then combines a robust graph autoencoder with contrastive learning to learn batch-corrected, spatially informed embeddings for identifying spatial domains across the tissue sections. These embeddings are then coupled with GWAS statistics to map trait-associated signals onto the integrated tissue sub-structure landscapes spanning different sections, developmental stages, and disease conditions. STUltra substantially outperforms competing methods, especially being much faster, on diverse integration benchmarks, enabling efficient processing of datasets containing millions of spots. As a case study on a mouse embryo dataset, STUltra recovered continuous developmental programs and mapped congenital heart disease signals beyond cardiac regions to broader programs including vascular and mesenchymal domains. On a mouse Alzheimer's disease dataset, STUltra prioritized localized AD-associated genetic signals to microglia and a disease-expanded C1QA/HEXB-high astrocyte state. Together, STUltra provides a scalable framework to bridge the gap between spatiotemporal tissue structures and complex disease traits, facilitating discoveries from previously unrecognized molecular and cellular architectures across diverse contexts.