Automated B-cell and plasma cell identification using unsupervised clustering by FlowSOM and an excel-based...

Area of Science:

• Immunology
• Computational Biology
• Biotechnology

Background:

• Manual plasma cell (PC) identification in multiparametric flow cytometry (MFC) is laborious and subjective, especially for rare cell populations.
• Existing AI methods like FlowSOM require expert interpretation, limiting standardization in research.
• There is a need for accessible, transparent tools to automate and standardize B-cell and PC identification in flow cytometry.

Purpose of the Study:

• To develop a lightweight, transparent, spreadsheet-based algorithm for classifying B-cells and PCs in flow cytometry.
• To integrate this algorithm with automated clustering outputs (FlowSOM) for standardized identification in research datasets.
• To create a tool that reduces operator dependency and improves inter-case harmonization.

Main Methods:

• Developed an Excel-based classifier ('PC Trainer Classifier') utilizing normalized marker intensities and a scoring system (PCscore, NEOscore).
• Integrated the classifier with FlowSOM clustering outputs from bone marrow aspirates stained with a standard PC screening tube.
• Trained the rule-based classifier on expert-assigned clusters and validated it on independent datasets.

Main Results:

• The B-cell classifier achieved high performance: Sensitivity 0.902, Specificity 0.984, Precision 0.885, Accuracy 0.973, and F1 0.893.
• Plasma cell identification showed moderate performance: Sensitivity 0.651, Specificity 0.828, Precision 0.458, Accuracy 0.796, and F1 0.537.
• The PC classification is suggested as a triage tool due to biological heterogeneity.

Conclusions:

• A transparent Excel-based classifier integrated with FlowSOM enables reproducible B-cell identification and structured PC classification in research flow cytometry.
• The tool offers interpretability, low cost, and portability, suiting resource-variable research settings.
• While B-cell classification is highly accurate, PC identification requires further optimization and validation, serving best as an augmented intelligence aid.