Extended Data Fig. 10: Cartoon of semi-synthetic benchmark construction presented in Figure 2.
From: Deep generative modeling of sample-level heterogeneity in single-cell genomics
Data from a PBMC dataset of 68K cells [29] are clustered using the Leiden algorithm. Then, for cluster A, synthetic DE effects are introduced by assigning subclusters of cells to different study subjects. For clusters B and C, different proportions of each cell type are sampled depending on the study subject, producing a ground truth DA effect.
