Figure 3.

Distribution of prediction errors in kinase-inhibitor interaction activity models. Shown are prediction errors in models using three different data modelling methods, namely: support vector machines (Panel A), partial least-squares projections to latent structures (Panel B), and k-nearest neighbour approach (Panel C). Prediction errors are estimated by outer-loop cross-validation, iteratively excluding 1/5 of the kinases in the data set. The histograms represent the absolute values of prediction errors (i.e. blue bars; labelling on the left side of panels); the cumulative plot of prediction errors is represented by red lines; labelling on the right side of panels).