Illustration of cross-predictive power for a number of models derived from public datasets. Trivial models ("All compounds are clean" and "All compounds are toxic") are compared to models developed in this study ("This study") and in reference  ("Ref"). Arrows indicate the direction of prediction. The percentage shown below each arrow is the percentage of correctly classified compounds: (true positives+ true negatives)/all. The toxicity cutoff of the Scripps dataset (b) was defined in ref resulting in 83% toxic compounds. Toxic and non-toxic sets are shown in red and green, respectively.