We propose a method to identify the correct binding mode of a ligand with a protein among multiple predicted docking poses. Our method consists of two steps. First, five independent MD simulations with different initial velocities are performed for each docking pose, in order to evaluate its stability. If the root-mean-square deviations (RMSDs) of heavy atoms from the docking pose are larger than a given threshold (2.0 Å) in all five parallel runs, the pose is filtered out and discarded. Then, we perform accurate all-atom binding free energy calculations for the residual poses only. The pose with the lowest binding free energy is identified as the correct pose. As a test case, we applied our method to a previously built cross-docking test set, which included 104 complex systems. We found that the present method could successfully identify the correct ligand binding mode for 72% (75/104) of the complexes for current test set. The possible reasons for the failure of the method in the other cases were investigated in detail, to enable future improvements.

我们提出了一种方法来在多个预测的对接姿势中识别配体与蛋白质的正确结合模式。我们的方法包括两个步骤。首先，对每个对接姿势进行了五个具有不同初始速度的独立 MD 模拟，以评估其稳定性。如果在所有五个平行运行中，重原子与对接姿势的均方根偏差 (RMSD) 都大于给定的阈值 (2.0 Å)，则该姿势将被过滤掉并丢弃。然后，我们仅对残余位姿进行精确的全原子结合自由能计算。具有最低结合自由能的姿势被识别为正确姿势。作为一个测试用例，我们将我们的方法应用于先前构建的交叉对接测试集，其中包括 104 个复杂系统。我们发现本方法可以成功地为当前测试集的 72% (75/104) 的复合物识别正确的配体结合模式。详细调查了其他情况下该方法失败的可能原因，以实现未来的改进。