基于广义回归神经网络的幼蛏毒性模型的建立与应用

Establishment and application of the toxicity on the larval Sinonvaculina constricta model based on GRNN algorithm

为建立芳烃化合物物化性质与缢蛏幼体(幼蛏)毒性间的非线性关系,本研究以实验获取的12种芳烃化合物对幼蛏96 h的毒性数据为基础,采用密度泛函理论(DFT)中的B3LYP方法,在6-311G^(**)基组上全优化计算12种芳烃化合物结构参数和热力学参数,采用广义回归神经网络(GRNN)和3层误差反向传播(BP)神经网络方法对实验数据进行回归拟合,模拟芳烃化合物的结构参数和热力学参数对幼蛏急性毒性的非线性关系。结果表明,GRNN的泛化结果和实测值的误差小于BP神经网络,预测准确性优于BP神经网络。t检验结果揭示建立的GRNN模型是可信的。相关性验证表明,GRNN模型具有良好的拟合度,在样本数据较少时也能很好的表达芳烃化合物对幼蛏急性毒性的非线性关系。利用GRNN模型预测未知芳烃化合物的实验结果,为芳烃化合物对幼蛏急性毒性关系的分析提供科学依据。

To investigate the nonlinear relationship between aromatic hydrocarbons and its toxicity on the larval Sinonvaculina constricta, optimized geometries of the 12 aromatic hydrocarbons were carried out at the B3LYP/6-311G＊＊ level with density functional theory（DFT）/B3LYP method. Afterwards, the structural and thermodynamic parameters obtained from the optimized geometries were taken as theoretical descriptors to establish the QSAR model with the generalized regression neural network（GRNN） method and three-layer backpropagation network＇s（BP） based on the experimental data of aromatic hydrocarbons on the larval Sinonvaculina constricta. The results show that the error of the actual data and generalization value of GRNN were smaller than BP neural network, and the accuracy of GRNN was superior to BP. T test showed that the GRNN model was credible. The verification of correlation showed that the GRNN model had the better fitting degree and nonlinear relationship of the aromatic hydrocarbons and its toxicity on the larval Sinonvaculina constricta. Based on the established GRNN model could be used to predicting the toxicity of the unknown aromatic hydrocarbons on the larval Sinonvaculina constricta, which is beneficial to analysis the toxicity of aromatic hydrocarbons on the larval Sinonvaculina constrict.