Evaluation on the PDB30 dataset.

Predictive performances of three freely and easily downloadable methods on Pdb30. The standard deviations were calculated over the same 100 bootstrap copies of the whole dataset. Given the huge size of the dataset, all differences (even if they are sometimes tiny) are statistically significant. Notice that (except for the AUC calculation), our method uses a classification threshold that was selected on the training dataset (Disorder723) so as to maximize the balanced accuracy, which explains its difference in (sensitivity, specificity) pattern, as compared to the other methods. Changing the threshold so as to yield a 94.7% specificity on Pdb30, would reduce its sensitivity to 58.8%.