基于GRNN模型的肼类化合物毒性参数构效分析

Structure-Activity Analysis of Toxicity Parameters of Hydrazine-Based Compounds Using GRNN Model

对30种肼类化合物的6个量子化学参数进行灰色关联度分析,利用广义回归神经网络(GRNN)模型对肼类化合物的毒性参数进行预测。通过与BP神经网络建立的模型进行比较,其平均预测误差分别为5.160%和12.667%。结果表明,所建立的GRNN有效解决了BP网络存在的过训练和过拟合问题,并且对肼类化合物毒性参数具有较好的预测效果。

The grey correlation analysis for 6 quantum chemistry parameters of 30 hydrazine-based compounds was carried out.The toxicity parameters of the hydrazine-based compounds were predicted by using the generalized regression neural network（GRNN） model.Comparing with the model established using BP neural network,their average prediction errors were 5.160% and 12.667% separately.The results show that the over-training and over-fitting problems existed in the BP network can be overcome using the established GRNN,and the better prediction effects for toxicity parameters of the hydrazine-based compounds can be obtained.