Cross-Modal Denoising and Integration of Spatial Multi-Omics Data with CANDIES.

Area of Science:

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

Background:

• Spatial multi-omics data provide cellular context but face challenges due to data quality variations and noise across modalities.
• Accurate integration and analysis of these diverse datasets are crucial for understanding tissue architecture and function.
• Existing methods struggle to effectively denoise and unify spatial multi-omics information.

Purpose of the Study:

• To introduce CANDIES, a novel computational framework for denoising and integrating spatial multi-omics data.
• To develop a unified joint representation that enhances downstream spatial analyses.
• To enable the interpretation of complex human traits within their spatial tissue context.

Main Methods:

• Utilized a conditional diffusion model for data denoising.
• Employed contrastive learning for data integration and representation learning.
• Evaluated performance on diverse synthetic and real spatial multi-omics datasets (MISAR-seq, CITE-seq, Mux-seq, ATAC-RNA-seq, Visium).

Main Results:

• CANDIES demonstrated superior performance in denoising and integrating spatial multi-omics data across multiple tissue types and modalities.
• The method generated a unified joint representation, significantly improving downstream tasks like spatial domain identification and trajectory reconstruction.
• CANDIES representations successfully integrated with genome-wide association studies (GWASs) to provide spatially resolved interpretations of complex human traits.

Conclusions:

• CANDIES offers a robust solution for enhancing the quality and utility of spatial multi-omics data.
• The framework facilitates deeper biological insights by unifying diverse molecular profiles in their spatial context.
• This approach opens new avenues for linking molecular mechanisms to complex traits at a spatially resolved level within tissues.