Fig. 2. Overview of the main results of this benchmark study.

a The DR methods reviewed in this study, their basic characteristics described in text, and their performances in terms of accuracy, scalability, stability, and utility. Results for the real and the simulation data were averaged for each of the four sub-categories. The DR methods were ranked based on the overall accuracy, averaging across all four sub-categories. All bars shown are calculated using ranks. Darker shaded bars within each column correspond to the better performance and therefore longer bars. Different colors are used to distinguish between different categories. Complementarity of the DR methods, evaluated on the real+simulated CyTOF datasets (b), the real CyTOF datasets (c), and the simulated datasets (d). We define complementarity as the likelihood of obtaining a top-performing method for a given dataset by choosing a specific DR method. The short vertical red lines represent the baseline. The red points represent the resulting likelihood of obtaining a top model by adding the best method that has not been previously added. The gray points show the resulting likelihoods if other methods are chosen instead of the remaining best one. Small random noises are added to the y-axis to differentiate the gray points. Source data are provided as a Source Data file.