Learning tissue representation by identification of persistent local patterns in spatial omics data

Area of Science:

• Computational biology
• Spatial transcriptomics
• Biomedical data analysis

Background:

• Spatial omics data offer molecular and structural insights into tissue heterogeneity.
• Analyzing spatial data can enhance patient stratification by linking clinical data to tissue characteristics.

Purpose of the Study:

• Introduce Kasumi, a novel method for identifying persistent, spatially localized neighborhood patterns of intra- and intercellular relationships.
• Demonstrate Kasumi's utility in translational tasks, specifically for cancer patient stratification based on disease progression and treatment response.

Main Methods:

• Developed Kasumi to detect persistent spatial patterns across samples and conditions.
• Applied Kasumi to spatial omics data from different experimental platforms.
• Evaluated Kasumi's performance against related approaches for patient stratification.

Main Results:

• Kasumi effectively represents tissues based on identified spatial patterns.
• The method outperforms existing approaches in stratifying cancer patients.
• Kasumi provides explanations for spatial coordination and relationships at cell-type or marker levels.
• Identified that persistent patterns vary in size and localized relationships correlate with poor outcomes.

Conclusions:

• Kasumi offers a robust method for analyzing spatial omics data and identifying clinically relevant tissue patterns.
• The discovered spatial relationships, even if localized, are crucial for predicting patient outcomes.
• Kasumi facilitates translational research by enabling refined tissue representations for improved clinical applications.