Figure 1

Schematic representation of predictive error distribution in DPRESS. The arrow weights indicate how much of the error made in predicting the response for the held-out compound (open symbol) is distributed among the compounds in the training set (solid symbols) when calculating the scaling factors γ_t. The data set is comprised of 13 observations in a two-dimensional descriptor space. Each panel represents one of the 13 separate analyses that make up the full leave-one-out (LOO) cross-validation run; only four of the 13 are shown.