A comparison framework and guideline of clustering methods for mass cytometry data

Area of Science:

• Biomedical Data Science
• Computational Biology
• Immunology

Background:

• Mass cytometry is increasingly used in medical research.
• Effective analysis of mass cytometry data relies on selecting appropriate clustering methods.
• Optimal method selection accelerates the identification of significant cell populations.

Purpose of the Study:

• To compare the performance of nine different clustering methods for mass cytometry data analysis.
• To evaluate unsupervised and semi-supervised clustering approaches.
• To provide guidelines for selecting suitable clustering tools.

Main Methods:

• Evaluated seven unsupervised (Accense, Xshift, PhenoGraph, FlowSOM, flowMeans, DEPECHE, kmeans) and two semi-supervised (Automated Cell-type Discovery and Classification, linear discriminant analysis (LDA)) methods.
• Assessed methods using precision (external), coherence (internal), and stability metrics across six benchmark datasets.
• Compared performance against random subsampling, varying sample sizes, and cluster numbers.

Main Results:

• Linear discriminant analysis (LDA) showed high precision but lower internal evaluation scores.
• PhenoGraph and FlowSOM outperformed other unsupervised methods in precision, coherence, and stability.
• PhenoGraph and Xshift were robust for sub-cluster detection; DEPECHE and FlowSOM tended to form meta-clusters.
• PhenoGraph, Xshift, and flowMeans performance varied with sample size, while FlowSOM remained stable.

Conclusions:

• A comprehensive evaluation of precision, coherence, stability, and resolution is crucial for selecting cytometry data analysis tools.
• Decision guidelines are provided to assist users in choosing the most appropriate clustering tools based on method characteristics.