DECIPHER for learning disentangled cellular embeddings in large-scale heterogeneous spatial omics data.

Area of Science:

• Computational biology
• Bioinformatics
• Spatial transcriptomics

Background:

• Current spatial modeling approaches obscure cellular function by failing to adequately represent the interplay between molecular identity and spatial context.
• Modeling large-scale heterogeneous spatial data requires efficient and effective computational methods.

Purpose of the Study:

• To introduce DECIPHER, a novel computational framework designed to disentangle intra-cellular and extra-cellular representations of cells.
• To enable effective and efficient in silico modeling of large-scale heterogeneous spatial data.
• To delineate cell-environment interactions across multiple scales.

Main Methods:

• DECIPHER utilizes a novel cross-scale contrast learning strategy.
• The method disentangles cellular representations into intra-cellular and extra-cellular components.
• The framework is designed for scalability to handle large spatial atlases.

Main Results:

• DECIPHER demonstrates superior performance compared to existing state-of-the-art methods.
• Systematic benchmarks and real-world case studies validate the method's effectiveness.
• The disentangled embeddings generated by DECIPHER enable accurate delineation of cell-environment interactions across various scales.

Conclusions:

• DECIPHER provides a scalable and effective solution for modeling complex spatial biological data.
• The framework enhances our understanding of cellular function by separating molecular and spatial contexts.
• DECIPHER's ability to handle millions of cells overcomes limitations of current spatial modeling techniques.