Figure 5.

Deconvolution of ligand-target interaction prediction by custom visual rendering of decision trees (DTs) comprising random forests. The visualizations provide pinpoint description of the criteria identified in the samples and features presented to the DT-building algorithm (samples and features provided by random selection with replacement, known as bagging). Here, trees were constructed using CATS2D ligand descriptors with (a) dipeptide or (b) tripeptide descriptors. Node color intensity indicates the amount of purity obtained in for a decision based. Yellow coloring denotes inhibitory activity, and cyan color denotes non-inhibitory activity. Individual MCC/F1 performance on the external set is show below each tree.