Figure 8

Estimating the predictive uncertainty profile for the CYP2D6 inhibition data set. The models shown have three hidden neurons and make use of 35 descriptors. Vertical dotted lines indicate voting thresholds. (A) Distribution of training pool predictions (blue), errors for model CYP2D6-1a (naïve voting threshold at 16.5; red), and errors for model CYP2D6-1b (threshold shifted to 27.5; green). Fitted beta binomial distributions are shown as dashed lines. (B) Calculated uncertainty profiles and distributions of training pool error rates for model CYP2D6-1a (naïve voting threshold at 16.5; red), and for model CYP2D6-1b (refined threshold at 27.5; green). Uncertainty profiles calculated from beta binomial distributions are shown as red and green dashed lines. The dashed black line indicates average class uncertainties. (C) Distribution of validation set predictions (blue), errors for model CYP2D6-1a (red), and errors for model CYP2D6-1b (green). Beta binomial distributions fitted to the training pool are shown as dashed lines. (D) Calculated uncertainty profiles and distributions of observed validation set error rates for model CYP2D6-1a (naïve voting threshold at 16.5; red), and errors for model CYP2D6-1b (refined threshold at 27.5; green). Uncertainty profiles calculated from beta binomial distributions fitted to the training pools are shown as red and green dashed lines, whereas the composite uncertainty profile is indicated by the black dashed line.