Figure 5.

Summary of the degree to which different experimental factors affected how well MoAs can be distinguished. (a) When using only the 9 µM concentration compared to all four concentrations, the AUC-ROC decreased by ≥ 0.05 for 32.1% (113) of the MoA/cell line pairs, whereas the AUC-ROC increased by ≥ 0.05 for only 3.4% (12) of the MoA/cell line pairs, resulting in a net AUC-ROC decrease of ≥ 0.05 for 28.7% (101) of the MoA/cell line pairs. Using only the 3 µM concentration resulted in a similar trend but compared to the “9 µM only”-case more MoA/cell line pairs benefited from the exclusion of the other concentrations. In each case all active treatments at the respective concentration(s) were considered for the AUC-ROC computation; for the “3 µM only”-case this means substantially fewer compounds than for the “9 µM only”- and “all concentrations”-cases. (b) When using only the information from the BFP channel compared to using the information from all three fluorescent channels, we observed a net AUC-ROC decrease of ≥ 0.05 for only 13.5% of the MoA/cell line pairs, indicating that most of the information is contained in the BFP channel alone. Adding the GFP channel reduces the net loss to 6.1%; adding the RFP channel reduces the net loss to 1.2%. We have not observed strong patterns as to which kind of markers (structural versus signalling / bright versus dim) are most beneficial for MoA distinction, or for which MoAs distinction is particularly sensitive across cell types to the exclusion of particular markers. In all cases, the AUC-ROC is based on the exact same set of active treatments (active calling based on the “all three channels”-signature). For the AUC-ROC analysis we repeated the mRMR feature selection for each considered combination of channels (rather than just dropping features related to the excluded channel(s) from the “all three channels”-signature) for an unbiased comparison between the different combinations of channels.