Studying the impact of batch effects on pattern discovery in mass cytometry data 2929

Abstract Description Mass cytometry (MC) measures over 50 proteins in single cells and enables profiling of immune cell composition. Machine learning and statistical methods identify cell types and interrogate their correlation with physiological conditions like disease or age. However, fluctuations in instrument readouts, called batch effects, can interfere with detecting critical biological variations. In a previous report on the EPIC data mining platform, we showed that batch-wise scaling effectively reduced batch effects in MC data. To visualize and quantify the impact of batch effects on sample stratification, we developed group similarity analysis (GSA). This technique combines dimension reduction (e.g., UMAP) with silhouette analysis to find patterns in summary statistics of clustering outputs. We used GSA to compare batch scaling with six other batch normalization algorithms. Using a dataset of 159 PBMC samples of healthy donors acquired in 35 MC runs, we show that batch normalization improves stratification into three age groups. A second case study of 153 samples from liver cancer patients demonstrates that batch normalization improves the segregation of cell composition according to tissue types (peripheral blood, liver). The conclusions from GSA agree with other methods for evaluating batch alignments, such as Earth Mover’s (EMD) or the coefficient of variation among bridging replicates. In summary, batch effect reduction promotes biological pattern discovery in cytometry data. Topic Categories Computational and Systems Immunology (COMP)