Isu Y, Nagashima U, Aoyama T, Hosoya H
Department of Information Sciences, Ochanomizu University, Tokyo, Japan.
J Chem Inf Comput Sci. 1996 Mar-Apr;36(2):286-93. doi: 10.1021/ci950108b.
A perceptron type neural network simulator for structure--activity correlation of molecules has been developed with two different learning methods, i.e., back-propagation and reconstruction methods. First by use of the back-propagation method the exo/endo branching of norbornane and norbornene derivatives was correctly predicted from the set of 13C NMR chemical shifts for various ring carbon atoms. Then the obtained correlation was analyzed by the reconstruction learning method. It was shown in this case that the NMR shifts for two carbon atoms out of seven have strong correlation with the exo/endo branching. Further, structure--activity correlation between the 13C NMR chemical shifts and carcinogenicity of 11 polycyclic aromatic hydrocarbons was also analyzed using the reconstruction method. It was demonstrated that neural network analysis is suitable for the elucidation of complicated structure--activity problems where many factors are nonlinearly entangled.
已开发出一种用于分子结构 - 活性相关性的感知器型神经网络模拟器，采用了两种不同的学习方法，即反向传播法和重构法。首先，通过反向传播法，从各种环碳原子的13C NMR化学位移数据集中正确预测了降冰片烷和降冰片烯衍生物的外型/内型分支。然后，用重构学习法对所得的相关性进行分析。在这种情况下表明，七个碳原子中的两个碳原子的NMR位移与外型/内型分支有很强的相关性。此外，还使用重构法分析了11种多环芳烃的13C NMR化学位移与致癌性之间的结构 - 活性相关性。结果表明，神经网络分析适用于阐明许多因素非线性纠缠在一起的复杂结构 - 活性问题。
