The overlap between predicted pockets and bound ligands in holo protein structures from the LigASite database.

The first column gives the fraction of proteins for which a method's predictions overlap a ligand. The second column (Prediction Vol.) lists the average volume of the predicted pockets for each protein, while the third column (Ligand Vol.) lists the average volume of ligands observed in the structure. The next columns give the average volumes of the Intersection and Union of the predictions and ligands and the Jaccard coefficient (Intersection/Union). ConCavity and Structure predict pockets of similar sizes---both use a similar pocket volume threshold---but ConCavity's predictions overlap more of the bound ligands. ConCavity's higher Jaccard coefficient demonstrates that it better manages the tradeoff between precision and recall.

第一列给出某方法的预测结果与配体重合的蛋白质占比。第二列「预测体积（Prediction Vol.）」列出各蛋白质的预测结合口袋的平均体积，第三列「配体体积（Ligand Vol.）」则列出结构中观测到的配体的平均体积。后续列依次给出预测结果与配体的交集、并集的平均体积，以及雅卡尔系数（Jaccard coefficient，即交集与并集的比值）。ConCavity与Structure预测的结合口袋尺寸相近——二者均采用了相似的口袋体积阈值——但ConCavity的预测与结合配体重合的比例更高。ConCavity更高的雅卡尔系数表明，其在精准率与召回率的权衡上表现更优。