A challenge in CMap interpretation is that the analysis returns a rank-ordered list of connections, leaving the user to extract biological meaning from the list. We reasoned that while any given member of an MOA class would likely have a multitude of targets, integrating signatures across several examples of an MOA class would sharpen the on-target signal, while diminishing off-target effects. We codified this by identifying compounds that share MOA and by identifying genetic perturbagens belonging to the same gene family or were targeted by the same compounds. These perturbagen classes (PCLs) were then further refined by excluding compounds that failed to connect with their cognate class members based on L1000 connectivity analysis (see STAR Methods and Figure 4A). This yielded 171 high confidence PCLs (Table S7).
