Novel FlowSOM Meta-clustering Method leveraging Insights from Cell Biology 2307173

Abstract

Abstract Introduction FlowSOM is widely used for cytometry analysis due to the expansion of simultaneously detectable fluorochromes to over 40. FlowSOM clusters cytometry data, followed by consensus meta-clustering that integrates similar clusters into meta-clusters. However, consensus meta-clustering has some intricacies, such as setting the number of meta-clusters as a hyperparameter, annotating the obtained meta-clusters manually, and reassigning meta-clusters when the results do not align with desired analytical outcomes. Methods We developed a novel meta-clustering algorithm, which leverages existing biological knowledge to address the areas of improvement. After SOM clustering is performed, this algorithm utilizes a Marker Expression Table (MET) that links each cell type to its expected marker expression pattern. In the proposed method, phenotypes are assigned to SOM clusters computationally. Each cluster is annotated with cell types by matching the expression patterns in the MET with the phenotypes of the clusters. Finally, clusters are integrated into meta-clusters according to their annotated cell types. Results We conducted experiments comparing the consensus meta-clustering method and our proposed method using 23-color and 45-color human PBMC datasets acquired on Sony’s ID7000TM Spectral Cell Analyzer. Our method showed a further improvement in clustering quality, with the Macro F1 scores increasing by more than 0.1 for both datasets compared to the consensus meta-clustering method. Conclusion We developed a novel FlowSOM meta-clustering algorithm, which leverages biological information. Our method demonstrated improved accuracy, as evidenced by higher Macro F1 scores compared to consensus meta-clustering. The proposed method reduces the need to specify the number of meta-clusters, improves cluster interpretability by assigning cell type names, and can substantially decrease the amount of manual adjustment required to achieve the desired meta-clustering results, thanks to its high clustering quality. Funding Source n/a Topic Categories Computational and Systems Immunology (COMP)