Vector-guided graph learning for spatial multi-slice multi-omics alignment.

Area of Science:

• Computational biology
• Spatial transcriptomics
• Multi-omics data analysis

Background:

• Spatial mapping of multi-omics data is crucial for understanding cellular components across spatiotemporal axes.
• Conventional graph neural networks fail to account for directional and angular influences, limiting analysis of complex spatial structures.

Purpose of the Study:

• To develop a novel vector-guided lightweight graph model, stLVG, for enhanced spatial mapping, label transfer, and niche identification in multi-slice multi-omics datasets.
• To address limitations of existing methods by incorporating directional and angular information in neighbor aggregation.

Main Methods:

• stLVG employs adversarial learning with distance- and direction-informed weights to aggregate neighbor information.
• It learns two distinct shared feature spaces across slices.
• A multi-view contrastive learning framework integrates these features for comprehensive analysis.

Main Results:

• stLVG demonstrates superior performance across various technologies, modalities, and resolutions.
• The model accurately delineates critical regions, such as tumor edges in breast cancer samples.
• Achieves efficient execution on standard hardware, ensuring scalability for large-scale spatial omics studies.

Conclusions:

• stLVG offers a significant advancement in analyzing complex spatial structures within multi-omics data.
• Its efficiency and accuracy make it a valuable tool for spatial mapping, label transfer, and niche identification.
• The model's ability to incorporate spatial context enhances the dissection of cellular heterogeneity.