stMixer for Scalable Mosaic Integration and Label Transfer in Spatial Histology and Multi-Omics.

Area of Science:

• Computational biology
• Bioinformatics
• Spatial transcriptomics

Background:

• Integrating spatial histology with multi-omics data is crucial for high-resolution tissue analysis.
• Incomplete data overlap across sections limits coherent integration and cross-condition studies.

Purpose of the Study:

• To present stMixer, an unsupervised framework for integrating multi-modal spatial data.
• To enable biologically coherent integration and anatomically faithful label propagation across tissue sections.

Main Methods:

• Employs self-looped cross-attention for joint encoding of histological, molecular, and spatial features.
• Utilizes multi-modal metric learning for cross-section integration.
• Applies a graph-guided, cluster-level voting algorithm for label propagation.

Main Results:

• Demonstrates superior scalability and accuracy in dimensionality reduction, batch correction, and label transfer across six spatial modalities.
• Accommodates large, heterogeneous datasets across tissues, species, and technologies.
• Uncovers transient biological states and resolves fine-grained microstructures, correcting mis-annotations.

Conclusions:

• stMixer provides a versatile framework for advanced spatial multi-omics data integration.
• The method enhances biological discovery by enabling accurate label transfer and analysis of complex tissue architectures.
• stMixer facilitates mosaic integration, pseudo-time inference, and cross-tissue knowledge transfer.