Segmentation aware probabilistic phenotyping of single-cell spatial protein expression data.

Area of Science:

• Biotechnology
• Computational Biology
• Immunology

Background:

• Spatial protein expression technologies offer deep insights into cellular organization.
• Accurate single-cell segmentation is a critical bottleneck for analyzing multiplexed imaging data.
• Existing methods struggle with segmentation errors, impacting phenotype and cluster interpretation.

Purpose of the Study:

• To introduce STARLING, a probabilistic machine learning model for robust cell population quantification from spatial protein expression data.
• To address and account for segmentation errors inherent in spatial biology data analysis.
• To improve the accuracy and reliability of cellular phenotype and cluster identification.

Main Methods:

• Development of STARLING, a probabilistic machine learning model.
• Creation of a benchmarking workflow using multiplexed imaging data of cell line standards.
• Generation of spatial expression data from human tonsil tissue.
• Establishment of a biological plausibility score for cellular phenotypes.

Main Results:

• STARLING effectively quantifies cell populations while accounting for segmentation errors.
• Benchmarking demonstrates STARLING's superior performance in controlled and complex biological samples.
• Analysis of human tonsil tissue reveals novel cellular states and quantifies heterogeneity.
• STARLING identifies cell types missed by other methods.

Conclusions:

• STARLING provides a robust solution for analyzing spatial protein expression data, overcoming segmentation challenges.
• The model enhances the discovery and quantification of cellular heterogeneity in complex tissues.
• This work significantly advances the field of spatial biology and its applications in disease research.