Mosaic integration of spatial multi-omics with SpaMosaic

Abstract With the advent of spatial multi-omics, we can mosaic integrate diverse datasets with partially overlapping modalities to construct consensus multi-modal spatial atlases of the source tissue. SpaMosaic is a spatial multi-omics mosaic integration tool that employs contrastive learning and graph neural networks to construct a modality-agnostic and batch-corrected latent space suited for analyses like spatial domain identification and imputing missing omes. Using simulated and experimentally acquired datasets, we benchmarked SpaMosaic against single-cell multi-omics mosaic integration methods. The experimental data encompassed RNA and protein abundance, chromatin accessibility or histone modifications, acquired from brain, embryo, tonsil, and lymph node tissues. SpaMosaic achieved superior performance over existing methods in identifying known spatial domains with enhanced resolution and clarity while reducing noise and batch effects. It also ranked top for modality alignment, enabling seamless diagonal integration without the need for image registration. After integration, SpaMosaic can also impute missing modalities. With a mosaic set of mouse brain data with RNA and different epigenomic modalities, we integrated and imputed the missing omics. There we found the imputed gene activity scores of activating and silencing histone marks show the correct correlation with RNA. Moreover, we uncovered more region-specific genes and pathways showing the correct transcriptome-epigenome correlations in the imputed histone modification than in the measured chromatin accessibility modalities. Lastly, SpaMosaic’s imputation also allows the inference of relationships between different modalities without requiring co-profiling from the same section.