Fig. 2

Analysis of the link between chemical similarity and classification performance. a Classification rate during LOO-CV vs. mean 5 NN Tanimoto compound similarity. A clear link between correct classification rate (%) and chemical similarity is observed in the DILIrank (−vLessConcern) dataset. The only exception from this was the first bin, which was defined by only a single compound (n = number of correctly classified compounds) and hence not a representative rate), and indicated that external test set compounds that are more structurally similar to the training set were predicted better. b Distribution of the mean 5 nearest neighbour inter-similarities between the DILIrank (−vLessConcern) training dataset and the corresponding test sets. It was found that the 55 FDA validation set compounds had comparable structural similarity to the training set (orange) as the compounds within the external test set (blue). Both histograms are left-closed