Riemannian Metric Learning for Alignment of Spatial Multiomics.

Area of Science:

• Computational Biology
• Bioinformatics
• Spatial Omics

Background:

• Spatial technologies now profile multiple molecular layers (e.g., transcriptome, epigenome, proteome) within single tissue samples.
• Integrating these diverse spatial datasets presents a significant challenge due to heterogeneous feature spaces.
• Existing methods often struggle with aligning arbitrary modalities while preserving both spatial and feature information.

Purpose of the Study:

• To develop a robust framework for aligning spatial multiomics data across different modalities.
• To address the limitations of current techniques in integrating heterogeneous spatial feature spaces.
• To enable more comprehensive analysis of tissue architecture and cellular interactions.

Main Methods:

• Introduction of Manifold Gromov-Wasserstein (MGW), a metric-learning framework.
• Exploitation of the product structure of spatial multiomics data.
• Inference of modality-specific Riemannian pull-back metrics using neural fields.
• Alignment of Riemannian distances via Gromov-Wasserstein optimal transport for a hyperparameter-free cost.

Main Results:

• MGW successfully aligns spatial data from diverse modalities, including transcriptomics, metabolomics, and imaging.
• Demonstrated effectiveness on multiple datasets: mouse embryo spatiotemporal transcriptomics, colorectal cancer spatial transcriptomics, and human tissue spatial multiomics.
• MGW recovers biologically meaningful correspondences and spatially coherent tissue structures.

Conclusions:

• MGW provides a powerful and flexible approach for spatial multiomics data integration.
• The framework outperforms existing optimal transport (OT) and non-OT based methods for multi-modal spatial alignment.
• MGW facilitates deeper insights into tissue organization and cellular heterogeneity across different biological contexts.