Comparison of clustering methods for high-dimensional single-cell flow and mass cytometry data

Area of Science:

• Immunology
• Computational Biology
• Biotechnology

Background:

• High-dimensional flow cytometry and mass cytometry (CyTOF) generate complex datasets.
• Manual gating is inefficient and unreliable for analyzing these large, high-dimensional datasets.
• Automated clustering methods are crucial for unsupervised analysis and cell population identification.

Purpose of the Study:

• To perform an up-to-date, extensible performance comparison of automated clustering methods for high-dimensional cytometry data.
• To evaluate methods using publicly available immunology datasets with expert-defined cell populations as ground truth.
• To identify robust and efficient clustering algorithms for flow cytometry and CyTOF data analysis.

Main Methods:

• Evaluated multiple unsupervised clustering algorithms on diverse, publicly available immunology datasets.
• Used expert manual gating as the reference standard for cell population identification.
• Compared methods based on accuracy, speed, and suitability for high-dimensional data.

Main Results:

• Several methods, including FlowSOM, X-shift, PhenoGraph, Rclusterpp, and flowMeans, demonstrated strong performance.
• FlowSOM exhibited particularly fast runtimes, making it suitable for interactive, exploratory analysis.
• The study provides a benchmark for existing methods and a framework for future comparisons.

Conclusions:

• Automated clustering methods are effective for analyzing high-dimensional cytometry data.
• FlowSOM is a highly efficient option for large-scale exploratory analysis.
• The provided resources (scripts and data) facilitate reproducible research and extension of these comparisons.