基于GLM模型和神经网络研究芳烃化合物对藻类毒性

工业的快速发展和溢油事故的频繁发生所产生的大量芳烃化合物正威胁海洋生态系统的健康,建立芳烃化合物物化性质与小球藻急性毒性间的非线性模型,是预测未知芳烃化合物对藻类毒性的主要手段之一.本研究以实验获取的25种芳烃化合物对小球藻96h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25种芳烃化合物结构参数和热力学参数,通过逐步广义线性回归(GLM)、小波神经网络(WNN)和T-S模糊神经网络(T-SFNN)等方法,对芳烃化合物物化性质和藻类抑制毒性的非线性关系进行拟合和逼近.F检验表明,逐步GLM方程是显著的(p<0.001).配对t检验表明,GLM、WNN、T-SFNN3种模型都是可信的;决定系数(R2>0.96)表明3个模型具有很高的精确性.上述结果证明本文建立的模型具有良好的拟合度和解释能力,可以预测未知芳烃化合物对小球藻的急性毒性.模型误差计算结果表明,WNN的精度最高(mse=0.0076,mae=0.0533),采用WNN方法进行建模,预测未知芳烃化合物对小球藻的急性毒性是最合适的.

Toxicity of aromatic hydrocarbons on algal based on GLM model and neural networks

The health of marine ecosystems are affected by aromatic hydrocarbons generated by the rapid industrial development and frequent ocourrence of oil spill accident.The establishment of nonlinear relationship between the physical and chemical properties of aromatic hydrocarbons and the inhibition toxicity of Chlorella vulgaris was one of the main method to predict the unknown aromatic hydrocarbons toxicity.Based on the 96 hr-EC50 of the inhibition toxicity of aromatic hydrocarbons to Chlorella vulgaris and the optimized geometries of aromatic hydrocarbons carried out at the B3LYP/6-311G* * level by density functional theory(DFT) calculation,the nonlinear relationship between aromatic hydrocarbons and the inhibition activity were fit and approximated with General Linear Model(GLM),Wavelet Neural Networks(WNN) and Takagi-sugeno Fuzzy Neural Networks(T-SFNN) models.The significance of GLM equation was confirmed by F test(p<0.001),and the three models,GLM,T-SFNN and WNN,were reliable by matched pair t test with high precision by the determinative coefficient(R2>0.96).The results showed that three models could be used to predict the unknown aromatic hydrocarbons toxicity well.The WNN model with mean square error(mse) 0.0076 and the mean absolute error(mae) 0.0533 is the best choice for forecasting the unknown aromatic hydrocarbons toxicity.