gms | German Medical Science

Background: Compared to serum or transcriptome analysis, flow cytometry enables detection of biomarkers immediately related to its cellular origin without effects of dilution by secretion and spreading into the system or by lysis of cell preparations with cells at different levels of stimulation or clonality. This promises a new level of diagnostic power, the dominant challenges in diagnostic biomarker discovery. Advancing into new technologies of multiple color flow cytometry or even mass cytometry offers the possibility of screening for dozens of antigens on a single cell. These technologies capture highly complex datasets, which require new bioinformatic tools for automated, comprehensive and user-independent analysis.

Methods: Methods include cell-clustering with Expectation Maximization (EM)-iterations in a t-mixture model to identify populations in each sample. The integrated classification likelihood (ICL) criterion stabilizes the number of reasonable cell-clusters in each sample. No compensation process is required. Comparing two samples requires meta-clustering of cell-clusters from each sample. Assuming Gaussian distributions, the Bhattacharyya coefficients of cell-clusters and meta-clusters reveal the probability measure for grouping of identical populations.

Results: Investigating paired samples of peripheral blood (PB) and synovial fluid (SF) from patients with rheumatoid arthritis with the new algorithm revealed up to 100 cell clusters when staining with 10 different parameters. Clustering of up to 106 cell events takes less than 1 hour on a standard PC and about 5-10 min on a specially equiped server. In both types of samples, all major leukocyte populations were quantified, including neutrophils, eosinophils, T-cells and their sub-populations, monocytes, B-cells, and NK-cells. Reasonable changes between the PB and SF revealed changes in cell frequencies (e.g. B-cells) up to complete depletion (e.g. naïve CD4-T-cells) of populations. Conventional manual analysis confirmed these quantities with ≤3% deviation in populations >50,000 events and ≤10% in populations <5,000 events. The algorithm was also able to dissect profiles generated with of CyTOF technology using more than 26 parameters.

Conclusion: In conclusion, the results give a reasonable starting point to face the next field of research for marker detection and prediction analysis. The approach is not only applicable to fluorescence-based flow data but also mass spectrometry-based cytometry with many more parameters per cell.