A graph neural network-based spatial multi-omics data integration method for deciphering spatial domains

Recent advancements of spatial sequencing technologies enable measurements of transcriptomic and epigenomic profiles within the same tissue slice, providing an unprecedented opportunity to understand cellular microenvironments. However, effective approaches for the integrative analysis of such spatial multi-omics data are lacking. Here, we propose SpaMI, a graph neural network-based model which extract features by contrastive learning strategy for each omics and integrate different omics by an attention mechanism to integrate spatial multi-omics data. We applied SpaMI to both simulated data and three real spatial multi-omics datasets derived from the same tissue slices, including spatial epigenome–transcriptome and transcriptome–proteome data. By comparing SpaMI with the state-of-the-art methods on simulation and real datasets, we demonstrate the superior performance of SpaMI in identifying spatial domain and data denoising.Author summary: The rapid advancement of spatial sequencing technologies now enables the simultaneous measurement of multiple omic features within the same spot. However, high levels of data noise and inherent sparsity pose significant challenges to the effective integration of such spatial multi-omics data. To address this, we present SpaMI, a deep learning-based model designed to integrate spatial multi-omics data and produce a complementary, comprehensive representation. SpaMI incorporates a contrastive learning strategy, an attention mechanism, and cosine similarity regularization. The contrastive learning component facilitates more effective learning of modality-specific embeddings while mitigating the influence of noise, and the attention mechanism adaptively aggregates embeddings across different modalities. The resulting cell representation offers powerful support for downstream analyses such as cell clustering, spatial domain identification, and differential expression detection.