Mosaic integration of spatial multiomic data based on hierarchical graph contrastive learning with SpatialMOSI

Area of Science:

• Computational Biology
• Spatial Omics
• 3D Tissue Reconstruction

Background:

• Spatial omics technologies enable multimodal molecular profiling in situ.
• Integrating multislice spatial omics data is key for 3D tissue landscape reconstruction.
• Existing methods lack a unified framework for simultaneous cross-omic and cross-slice integration.

Purpose of the Study:

• To present SpatialMOSI, a unified framework for mosaic integration of multimodal spatial omics data.
• To address cross-modality and cross-section variations concurrently.
• To enable comprehensive 3D tissue reconstruction from diverse spatial omics datasets.

Main Methods:

• SpatialMOSI utilizes a hierarchical graph contrastive learning (HiGCL) strategy.
• Integrative objectives include cross-omic alignment, cross-slice batch correction, and spatial microenvironment preservation.
• The framework operates on modality-specific latent representations with decoding reconstruction for feature fidelity.

Main Results:

• SpatialMOSI demonstrates versatility across multiple biological systems.
• Accurate identification of spatially conserved domains and imputation of missing omic layers.
• Revealed insights into B cell dynamics, tumor-immune interactions, and embryonic development.

Conclusions:

• SpatialMOSI provides a unified computational framework for integrating multimodal spatial omics data.
• Enables simultaneous resolution of cross-omic and cross-slice variations.
• Establishes a foundation for constructing integrative 3D molecular atlases.