Extended Data Fig. 4. In silico downsampling experiments.

We performed in silico cell mixing of RPE-1 and HG01573 cells to simulate application of scNOVA to different cell fractions (CFs). In this analysis six different CF ranges were considered (20, 10, 5, 3.3, 2, and 1.3). For each in silico cell mixing experiment, a total of 150 single cells were randomly subsampled for the major pseudo-clone (containing RPE-1 cells) and the minor pseudo-clone (HG01573 cells), by controlling the minor pseudo-clone CF at 20, 10, 5, 3.3, 2, and 1.3%, respectively. AUC, area under the curve. DEGs, differentially expressed genes. For each CF, we performed random subsampling of single-cell libraries 10 times, and depicted the respective mean AUC in the plot. Two different analysis modes - default (dashed lines, CNN with negative binomial generalized linear model), and alternative (solid lines, CNN with PLS-DA) are depicted. When the CF is larger than 10%, the default mode performs better, whereas for CFs smaller than 10%, the alternative mode outperforms the default mode.