Supplementary Figure 8: Comparison between NicheNet, one-vs-one-vs-one models and leave-one-in models in target gene and ligand activity prediction performance.

To investigate the importance of the integration of multiple complementary data sources, we compared the performance of NicheNet to ‘one-vs-one-vs-one’ and ‘leave-one-in’ models. In one-vs-one-vs-one models, ligand-target regulatory potential scores were calculated after using only one comprehensive ligand-receptor, one signaling and one gene regulatory database (n = 280 one-vs-one-vs-one models in total). In ligand-signaling leave-one-in models, one ligand-signaling data source was used with all gene regulatory data sources; in gene regulatory leave-one-in models the opposite (n = 37 ligand-signaling and n = 20 gene regulatory leave-one-in models in total). For each separate one-vs-one-vs-one and leave-one-in model, we calculated the median target gene and ligand activity prediction performances over all ligand treatment datasets (111 datasets for 51 unique ligands) and compared to the median performances of the complete NicheNet model. AUPR (corrected): area under the precision-recall curve, corrected for random prediction; AUROC: area under the receiver operating characteristic curve. Red dashed line: performance of random guessing.