Back-propagation network improved by conjugate gradient based on genetic algorithm in QSAR study on endocrine disrupting chemicals  

作　　者：JI Li WANG XiaoDong YANG XuShu LIU ShuShen WANG LianSheng 

机构地区：[1]State Key Laboratory of Pollution Control and Resources Reuse, School of Environment, Nanjing University, Nanjing 210093, China [2]Key Laboratory of Yangtze River Water Environment, Ministry of Education, College of Environmental Science and Engineering, Tongji University, Shanghai 200092, China

出　　处：《Chinese Science Bulletin》2008年第1期33-39,共7页

基　　金：the National Natural Science Foundation of China (Grant No. 20507008);the National Natural Science Foundation Key Project of China (Grant No. 20737001);the Natural Science Foundation of Jiangsu Province,China (Grant No. BK200418);the National Basic Research Program of China (973 Program) (Grant No. 2003CB415002)

摘　　要：Since the complexity and structural diversity of man-made compounds are considered, quantitative structure-activity relationships (QSARs)-based fast screening approaches are urgently needed for the assessment of the potential risk of endocrine disrupting chemicals (EDCs). The artificial neural net-works (ANN) are capable of recognizing highly nonlinear relationships, so it will have a bright applica-tion prospect in building high-quality QSAR models. As a popular supervised training algorithm in ANN, back-propagation (BP) converges slowly and immerses in vibration frequently. In this paper, a research strategy that BP neural network was improved by conjugate gradient (CG) algorithm with a variable selection method based on genetic algorithm was applied to investigate the QSAR of EDCs. This re-sulted in a robust and highly predictive ANN model with R2 of 0.845 for the training set, q2pred of 0.81 and root-mean-square error (RMSE) of 0.688 for the test set. The result shows that our method can provide a feasible and practical tool for the rapid screening of the estrogen activity of organic compounds.Since the complexity and structural diversity of man-made compounds are considered, quantitative structure-activity relationships （QSARs）-based fast screening approaches are urgently needed for the assessment of the potential risk of endocrine disrupting chemicals （EDCs）. The artificial neural networks （ANN） are capable of recognizing highly nonlinear relationships, so it will have a bright application prospect in building high-quality QSAR models. As a popular supervised training algorithm in ANN, back-propagation （BP） converges slowly and immerses in vibration frequently. In this paper, a research strategy that BP neural network was improved by conjugate gradient （CG） algorithm with a variable selection method based on genetic algorithm was applied to investigate the QSAR of EDCs. This resulted in a robust and highly predictive ANN model with R2 of 0.845 for the training set, q^2 pred of 0.81 and root-mean-square error （RMSE） of 0.688 for the test set. The result shows that our method can provide a feasible and practical tool for the rapid screening of the estrogen activity of organic compounds.

关 键 词：化学药物 内分泌 人造神经网络 遗传算法 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]