Supplementary Figure 13: Popularity bias of NicheNet, CCCExplorer and IPA Upstream Regulator Analysis in ligand activity prediction performance.

To analyze popularity bias in ligand activity prediction performance, we assessed performance after iteratively leaving out the ligand treatment datasets profiling the transcriptional response to the top n most popular ligands. Following ligand activity prediction evaluation metrics were calculated: AUPR (corrected): area under the precision-recall curve, corrected for random prediction; AUROC: area under the receiver operating characteristic curve. To analyze popularity bias and compare between NicheNet, CCCExplorer and IPA Upstream Regulator Analysis, we assessed the relation between popularity of ligands and median performance when datasets of these ligands are still included (n = 51 different ligands). The shown smoothing lines are the result of fitting a linear regression model to analyze this relation. The accompanying shaded region indicates the 95% confidence interval. The popularity of ligands is determined by the number of studies in PubMed in which a ligand is described.