Tensor decomposition to identify context-aware spatial neighborhoods in the tumor microenvironment

Area of Science:

• Computational Biology
• Cancer Research
• Bioinformatics

Background:

• The tumor microenvironment (TME) comprises diverse cells with crucial roles in tumor progression.
• Multiplex tissue imaging reveals cell types via biomarkers at single-cell resolution.
• Current image analysis tools quantify interactions per image, limiting scalability.

Purpose of the Study:

• To develop a scalable method for analyzing spatial heterogeneity of cell type distributions across multiple samples.
• To apply tensor decomposition to multiplex imaging data for comprehensive TME analysis.
• To identify cellular neighborhoods and their association with clinical outcomes in colorectal cancer.

Main Methods:

• Utilized a scalable canonical polyadic (CP) tensor decomposition method.
• Applied the method to a public colorectal cancer (CRC) dataset (56 markers, 35 patients, 136 images).
• Analyzed tensor decomposition factors to identify cellular neighborhoods and their variations.

Main Results:

• Identified tensor components corresponding to key cellular neighborhoods (e.g., bulk tumor, tumor edge).
• Wilcoxon rank sum tests revealed significant differences in components between Crohn's-like reaction (CLR) and diffuse inflammatory infiltration (DII) groups, noting granulocyte enrichment in DII.
• Survival analysis linked specific components to high tumor burden and poor patient prognosis.

Conclusions:

• Scalable CP tensor decomposition effectively probes spatial heterogeneity in the TME across multiple samples.
• The method can identify clinically relevant cellular neighborhoods and their associations with disease subtypes and patient outcomes.
• This approach offers a powerful tool for analyzing complex multiplex imaging data in cancer research.