神经网络法应用于酚类化合物对青海弧菌毒性的预测

Prediction of the Toxicity of Phenol Derivatives to Vibrio-qinghaiensis by Neural Network Method

青海弧菌对有毒酚类化合物具有强烈的敏感性,为建立酚类衍生物对青海弧菌毒性的定量结构-活性相关性(QSAR)模型,分析了16种酚类衍生物的分子结构与对青海弧菌毒性之间的相关关系,计算了酚类衍生物的分子连接性指数和分子形状指数,并优化筛选了分子连接性指数的1阶路径指数(~1χ)和分子形状指数的2阶特征指数(K_2)及1阶和2阶指数乘积值(K_4),用这3种指数与对青海弧菌的毒性进行多元回归分析,多元回归方程的决定系数R^2=0.971。为进一步提高预测精度,将这3种分子结构参数作为神经网络的输入变量,毒性值作为输出变量,采用3:2:1的网络结构,通过反向传播(BP)神经网络法获得满意的QSAR预测模型,总的相关系数r为0.996,计算得到的毒性预测值与实验值较为吻合,平均相对误差仅为1.98%,结果表明该模型具有良好的预测酚类衍生物毒性的能力,可以看出神经网络方法对酚类化合物发光菌毒性预测比多元线性回归方法的统计学意义更加明显。

Vibrio-qinghaiensis is very sensitive to toxic phenol derivatives. In order to establish QSAR of phenol derivatives to Vibrio-qinghaiensis, the relationship between molecular structure of 16 kinds of phenol derivatives and the toxicity to Vibrio-qinghaiensis was analyzed. Moreover, the molecular connectivity indices and spatial shape indices of phenol derivatives were calculated. The molecular connectivity index, 1χ, and spatial shape indices, K2 and K4, were screened. Then, multi-linear method was applied in analyzing the three indices and the toxicity to Vibrio-qinghaiensis. The determination coefficient R^2 was 0.971. In order to improve accuracy, the three indices were used as input variables of neural network and the toxicity was used as output variable, the 3：2：1 network structure was adopted and BP neural network method was used to establish a satisfying QSAR model. The total correlation coefficient r was 0.996. The predicted values were very close to experimental values, and the relative mean error was 1.98%, which showed that the model had good predictive ability of the toxicity of phenol derivatives. Futhermore, neural network method had more obivious statistical significance than multi-linear method.