LATTICE Integrates Multimodal Spatial Omics Data with Graph AI

A new research introduces LATTICE (Latent Alignment of Tissue-level and Transcriptomic Information for Cross-modal Embedding), a graph-based self-supervised learning framework designed for integrating multimodal spatial omics data. Spatially resolved omics studies are increasingly combining various assays, such as transcriptomics and epigenomics, but often lack unified analysis pipelines. LATTICE addresses this by learning comprehensive spot-level representations from harmonized features across multiple modalities. The framework integrates five aligned modality blocks per Visium spot: Visium RNA, scMultiome RNA, scMultiome ATAC, spatial ATAC, and spatial CUT&Tag. These diverse data types capture spatial gene expression, inferred regulatory activity, and in situ chromatin states. LATTICE constructs a spatial neighborhood graph and trains a TransformerConv encoder using objectives like masked reconstruction, cross-modal alignment, and spatial smoothness to create a unified lattice representation. Tested on a private melanoma cohort, LATTICE demonstrated stable optimization and reproducible embeddings. Integrating scMultiome RNA with Visium RNA alone substantially improved concordance with existing clusters and spatial contiguity. Adding further modalities enhanced spatial contiguity and a multimodal utility score, even if it sometimes diverged from purely RNA-derived labels, suggesting the embeddings captured richer chromatin and regulatory information. LATTICE offers a practical framework for multimodal spatial omics integration, while also highlighting the need for more robust external benchmarking.