Image-guided Spatial Omics Enhancement reveals Hidden Spatial Microstructures

Area of Science:

• Spatial omics
• Computational biology
• Deep learning

Background:

• Spatial omics advancement is limited by resolution gaps and data challenges.
• Current computational methods lack multi-modal flexibility and scalability for whole-tissue data.

Purpose of the Study:

• To introduce a deep learning framework for reconstructing high-fidelity spatial microstructures.
• To address the resolution gap and data sparsity in spatial omics.

Main Methods:

• Developed Bell, a modality-agnostic deep learning framework.
• Introduced mmBell, an extension for cross-modal integration of multi-omics data.
• Utilized an adaptive attention mechanism for dynamic data fusion.

Main Results:

• Bell and mmBell consistently outperform state-of-the-art methods.
• Achieved significant resolution enhancement and noise suppression.
• Validated across over 10 spatial platforms and 20 datasets.

Conclusions:

• Bell and mmBell offer a robust and scalable solution for spatial omics data.
• The framework enables deeper deciphering of complex spatial tissue organization.