SpaTranslator: A deep generative framework for universal spatial multi-omics cross-modality translation

Abstract Recent advances in spatial omics technologies have enabled the simultaneous analysis of multiple molecular patterns in tissue sections, offering unprecedented insights into cellular microenvironments. However, the high cost of measurements and the sparsity of data restrict the availability of paired spatial multi-omics datasets. Here, we present SpaTranslator, a deep generative framework that integrates graph neural networks with an adversarial variational generative model to fully capture spatial characteristics and enable effective cross-modality translation of spatial omics data, enabling simulation of paired spatial multi-omics data from single-omics measurements. Extensive experiments demonstrate that SpaTranslator consistently outperforms baseline methods in both clustering accuracy and biological coherence across various real-world scenarios, including spatial transcriptomics-epigenomics and spatial transcriptomics-proteomics translation tasks. Furthermore, SpaTranslator provides biologically meaningful insights through marker genes and proteins recovery, motif enrichment analysis, and gene regulation inference. Our work offers an effective and adaptable solution for spatial multi-omics cross-modality translation, supporting a broad range of biological and biomedical research. Competing Interest Statement The authors have declared no competing interest.