Docking Score Calculation Using Machine Learning with an Enhanced Inhibitor Database
Docking Score Calculation Using Machine Learning with an Enhanced Inhibitor Database
复制标题
使用机器学习和增强型抑制剂数据库计算对接分数
DOI：
10.1166/jmihi.2015.1503
复制
发表时间：
2015
影响因子：
--
通讯作者：
Hayato Ohwada and Shin Aoki
中科院分区：
文献类型：
--
作者：
Masato Okada;Tadasuke Ito;Hayato Ohwada and Shin Aoki
本文描述了一种用于计算与有用诱饵数据库对接得分的机器学习方法：增强型(DUD-E)。这个数据库既包括好的抑制剂(配体)，也包括差的抑制剂(诱饵)，使机器学习能够预测配体和相关诱骗的适当对接分数。这一特性使我们能够高精度地寻找新的候选缓蚀剂，并筛选出许多性能优异的化合物。该方法还可以应用于任何酶，而不需要使用酶的分子结构，在预测准确性和通用性方面都优于一些传统的对接软件工具。我们从DUD-E中选择了10种酶，并使用两个对接软件工具进行了比较研究。实验从2985个实际配体中预测了2869个配体，从8955个注册的诱饵中预测了8923个诱饵，从而获得了分类性能。通过各种阈值设置下的接收器工作特性(ROC)曲线，可以直观地看到如此出色的性能。
This paper describes a machine-learning method for docking score calculation with the Database of Useful Decoys: Enhanced (DUD-E). This database includes both good inhibitors (ligands) and poor inhibitors (decoys), allowing machine learning to predict appropriate docking scores of a ligand and the associated decoys. This property enables us to find new inhibitor candidates with high accuracy and to screen many compounds with excellent performance. The proposed method can also be applied to any enzymes without the use of the molecular structure of an enzyme, outperforming a number of traditional docking software tools in both predictive accuracy and generality. We selected 10 enzymes from DUD-E and conducted a comparative study using two docking software tools. The classification performance was obtained from an experiment where 2869 ligands were predicted from 2985 actual ligands, and 8923 decoys were predicted from 8955 registered decoys. Such excellent performance is visualized by Receiver Operating Characteristic (ROC) curves at various threshold settings.