Integrative deep learning of spatial multi-omics with SWITCH.

Area of Science:

• Computational Biology
• Genomics
• Neuroscience

Background:

• Spatial omics technologies enable spatially resolved biological measurements across multiple modalities.
• High costs limit the acquisition of co-profiled multimodal spatial omics data.
• Integrating unpaired spatial multi-omics data and performing cross-modal predictions is computationally challenging due to low signal-to-noise ratios.

Purpose of the Study:

• To develop a computational method for integrating unpaired spatial multi-omics data.
• To enable cross-modal predictions on single-modality spatial omics data.
• To improve the accuracy and resolution of spatial domain analysis.

Main Methods:

• Introduction of SWITCH (Spatially Weighted Multi-omics Integration and Cross-modal Translation with Cycle-mapping Harmonization), a deep generative model.
• Utilizing a cycle-mapping mechanism for dependable cross-modal translations without paired data.
• Employing cross-modal translations as pseudo-pairs to enhance data signals.

Main Results:

• SWITCH outperforms existing methods in spatial multi-omics integration accuracy.
• Achieved more precise spatial domain delineation, resolving brain cortical structures at higher resolution.
• Validated the reliability of cross-modal translations for downstream analyses.

Conclusions:

• SWITCH provides a robust framework for integrating unpaired spatial multi-omics data.
• The method facilitates advanced downstream analyses, including differential analysis, trajectory inference, and gene regulatory network inference.
• SWITCH enhances the utility of spatial omics data by enabling accurate cross-modal predictions and improved spatial resolution.