Spatially multimodal and multiscale network for representation learning from spatial multi-omics.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Spatial multi-omics techniques offer multilayered insights into tissue architecture, including gene expression and microenvironments.
• Integrating multimodal spatial omics data is crucial for downstream analyses like spatial domain clustering.
• Learning reliable latent representations from integrated spatial multi-omics data is essential.

Purpose of the Study:

• To develop an efficient method for learning latent representations from spatial multi-omics data.
• To improve the accuracy of spatial domain identification and tissue architecture analysis.

Main Methods:

• Developed Spatially Multimodal and Multiscale Network (SpaMM-Net).
• Utilized spatially-guided multiscale graph attention networks for integrating multimodal omics features.
• Incorporated multiscale feature integration to learn robust latent representations.

Main Results:

• SpaMM-Net demonstrated strong performance in spatial clustering tasks across multiple datasets.
• The method effectively deciphered both detailed regions and overall tissue architectures.
• Scale-wise weight analyses confirmed SpaMM-Net's effective use of multiscale spatial information for robust representations.

Conclusions:

• SpaMM-Net provides an efficient strategy for capturing and integrating latent representations from spatial multi-omics data.
• The developed tool is effective for spatial domain identification.
• The approach is extensible to other multi-omics modalities as spatial technologies advance.