Flexible Integration of Structurally Diverse Spatial Multi‐Omics With AgaeSMO

Spatial multi-omics (transcriptome, epigenome, proteome, metabolome, morphomics), provide near-panoramic omics profiles across tissues. However, cross-integration of structurally unpaired and underutilized morphomic datasets (e.g., Hematoxylin and eosin (H&E)-stained images), remains a bottleneck due to inherent differences in spot structures, feature dimensions, and cellular heterogeneities. Here, we present AgaeSMO, an adaptive graph auto-encoder for versatile cross-modality integration of both structurally paired and unpaired spatial multi-omics data. AgaeSMO encodes each omics modality from a cell type-aware KNN (k-nearest neighbor) feature graph and a spatial adjacency graph, while leverages GATs (Graph Attention Networks) as the backbone to fuse modalities by learning joint representations through an attention-weighted module. It specially fuses structurally unpaired modalities, which applies a morphological transformation strategy by designing a cross-guidance alignment among modalities through graph cross-attention learning. H&E images are treated as a standalone morphomic modality (rather than auxiliary features) is specially fused with other modalities through image patch embeddings to fully utilize cellular morphological heterogeneities. AgaeSMO is a powerful tool can effectively integrate diverse omics modalities with both paired and unpaired spot architectures. These unique capabilities enable precise spatial domain delineation, micro-environmental heterogeneity detection, compatible across broad platforms, that outperforming pre-existing tools. AgaeSMO offers a versatile, robust framework for integrative spatial multi-omics analysis and facilitates multimodal, biologically interpretable mining.