Spatial multi-omics imputation and embedding with SpaMIE.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Spatial multi-omics (SMO) data integration is challenged by missing modalities across tissue sections due to high costs and limited throughput of current platforms.
• Large-scale spatial atlases often feature heterogeneous modality coverage, with only a few sections fully profiled.
• This necessitates advanced computational methods to bridge data gaps and enable comprehensive analysis.

Purpose of the Study:

• To introduce SpaMIE, a deep graph neural network framework for effective multi-section integration of spatial multi-omics datasets.
• To address the challenge of systematic missing modalities in SMO data.
• To enable accurate cross-modal imputation and unified analysis of heterogeneous spatial data.

Main Methods:

• SpaMIE employs a two-stage approach: spatially informed cross-modal imputation and multi-section integration.
• The first stage infers missing modalities from mono-omics data using graph neural networks.
• The second stage integrates measured and imputed SMO profiles across sections to learn a unified embedding.

Main Results:

• SpaMIE demonstrates accurate cross-modal imputation of missing modalities.
• The framework achieves robust integration of spatial multi-omics data across multiple tissue sections.
• Benchmarking shows improved spatial domain identification and a flexible, scalable solution for SMO atlases.

Conclusions:

• SpaMIE provides a powerful computational solution for integrating spatial multi-omics data with missing modalities.
• The framework facilitates the construction and analysis of comprehensive SMO atlases.
• This approach enhances the utility of large-scale spatial datasets for biological discovery.