Comparison of pocket extraction methods.

For three types of grids (first column), we ran different pocket extraction algorithms (second column) and compared how well the pockets overlap bound ligands in holo PQS structures. The third column (“Prediction Vol.”) lists the average volume of all predicted pockets over each protein. For reference, the average volume of all ligands observed in the PQS files (“Ligand”) is 1977.2 Å3. The next two columns list the average volumes of the Intersection (Ligand Prediction) and Union (Ligand Prediction) of the Prediction and Ligand grids. Finally, the rightmost four columns list the average over-prediction factor (Prediction/Ligand), precision (Intersection/Prediction), recall (Intersection/Ligand), and Jaccard coefficient (Intersection/Union). For the last three columns, values range between zero and one, and higher values represent better performance. Comparing the average volume of the pockets predicted by each method, we see that Search's pockets are closest to the actual ligand volumes. Moreover, Search's high Jaccard coefficient for each grid type indicates that it provides the best tradeoff between recall and precision among the methods tested.

针对三类网格（第一列），我们运行了多种口袋提取算法（第二列），并对比了各算法提取的口袋与PQS结合配体结构中配体的重叠匹配程度。第三列（"Prediction Vol."）统计了每种蛋白质上所有预测口袋的平均体积。作为参考，PQS文件中观测到的所有配体的平均体积为1977.2 ų。后续两列分别列出了预测口袋与配体网格的交集（Ligand ∩ Prediction）及并集（Ligand ∪ Prediction）的平均体积。最后最右侧的四列依次给出了平均过预测因子（Prediction/Ligand）、精确率（Intersection/Prediction）、召回率（Intersection/Ligand）以及雅卡尔系数（Intersection/Union）。最后三列的取值范围均为0至1，数值越高代表算法性能越好。对比各方法预测得到的口袋平均体积，可以发现Search算法提取的口袋体积最接近真实配体体积。此外，针对每一类网格，Search算法均取得了较高的雅卡尔系数，这表明在本次测试的所有算法中，该算法在召回率与精确率之间实现了最优的平衡。