Fig. 6
From: Comparative evaluation of methods for the prediction of protein–ligand binding sites
Ligand binding site prediction benchmark at the pocket level. These curves correspond to the default predictions of the thirteen methods, indicated by (d) preceding their names. A Recall, percentage of observed sites that are correctly predicted by a method within the top-N+2 predictions; B recall using DCC = 12 Å but considering increasing rank thresholds, i.e., top-N, N+1, N+2, etc. ALL represents the maximum recall of a method, obtained by considering all predictions, regardless of their rank or score; C recall curve of top-N+2 predictions using Irel as a criterion; D ROC100 curve (cumulative TP against cumulative FP until 100 FP are reached); E precision curve for the top-1000 predictions of each method across the LIGYSIS dataset, Precision1K. Error bars represent 95% CI of the recall (A–C) and precision (E), which is 100 × proportion. Numbers at the right of the panels indicate groups or blocks of methods that perform similarly for each metric. Asterisks (*) indicate outlier methods, or methods that perform very differently than the rest