LATTICE: Graph Self-Supervised Learning for Multimodal Spatial Omics Integration

LATTICE (Latent Alignment of Tissue-level and Transcriptomic Information for Cross-modal Embedding) addresses the challenge of integrating heterogeneous spatial omics data. While many existing tools focus on transcriptomics alone, LATTICE constructs a spatial neighborhood graph and utilizes a TransformerConv encoder to process five distinct modality blocks: Visium RNA, scMultiome RNA, scMultiome ATAC gene scores, spatial ATAC, and spatial CUT&Tag. The model learns latent representations through three self-supervised objectives: masked feature reconstruction, cross-modal alignment, and spatial regularization. This design allows the framework to capture both molecular and spatial structure without requiring a separate alignment stage.

The researchers evaluated LATTICE using a modality ladder (M1–M5), where M1 represents Visium RNA only and M5 represents the full integration of all five modalities. The results show that adding scMultiome RNA to Visium RNA (M2) significantly improves concordance with standard RNA-derived clusters. However, as additional epigenomic modalities (spatial ATAC and CUT&Tag) are introduced in M4 and M5, the agreement with RNA-only reference labels decreases. Crucially, the authors demonstrate that this decrease does not indicate poor performance; rather, the embeddings are capturing chromatin and regulatory structures that transcriptomic data alone cannot see. These higher-level integrations (M4–M5) consistently yield higher spatial contiguity and a superior Multimodal Utility Score (MUS), suggesting that LATTICE successfully captures complex tissue organization.

As spatial omics technologies evolve to include multi-omic readouts, downstream analysis must move beyond single-modality pipelines. LATTICE provides a flexible, empirically grounded framework that can incorporate diverse data types into a unified representation. By demonstrating that multimodal integration can reveal spatial and regulatory patterns invisible to RNA-only methods, this work highlights the potential for graph-based self-supervised learning to serve as a standard paradigm for analyzing high-resolution spatial datasets.