Fig. 2

Schematic illustration of the benchmarking workflow. Using both synthetic and real single-cell datasets, six DA testing methods were evaluated under three configurations for the DA prediction task. A, B Single-cell RNA-seq and mass cytometry datasets were generated from patient samples, or synthetic datasets were generated using the packages dyntoy [20] or splatter [18]. C, D Next, we evaluated the six clustering-based [17] and clustering-free [1, 3,4,5,6] DA testing methods on datasets with different topologies, DA ratios, and technical biases such as batch effects. E Lastly, we compare the performance of the DA testing methods using both the AUROC and AUPRC scores