摘要
对结构参数采用主成分变换 ,再利用 BP人工神经网络 ,采用 L M算法作为迭代方法训练网络 ,预测检验集化合物的LD5 0 。结果显示 ,BP人工神经网络可以用于定量毒性构效关系研究 ,含隐层的 BP人工神经网络拟合能力明显优于传统方法 ,消除过度拟合后的多层 BP网络预测能力也好于传统方法 ,可以用于预测。
Objective Using BP Artificial Neural Network to study the Structure Activity relationship between aromatics compounds and rat LD 50 , improved precision of toxicity prediction. Methods Firstly, Principal Components Analysis was adopted, then used BP ANN net structure, and applied LM arithumetic as iteration method to train the network. Result We have discussed the relationship betwenn the structure parameter of 120 varieties of aromatics compound and rat LD 50 , and optimized the parameter design of the net to avoid over fitting. I found that three layer BP ANN which using log sigmoid function, (i.e.) f(x) =1/1(+exp( -x )) as network transfer function got better fitting power. When the number of the hidden layer node is 13, the sum square error is 0.36 which is far less than linear models. While the outer prediction precision of multiplayer BP ANN is higher than linear model in evidence, SSE=4.63. Conclusion We can consider that the classify power of multiplayer BP ANN is superior to linear nodels. Multilayer BP ANN can be use to predict toxicity of aromatics compounds, this method is better than traditional methods.
出处
《数理医药学杂志》
2001年第1期1-6,共6页
Journal of Mathematical Medicine
关键词
BP人工神经网络
LM算法
LD50
过度拟合
HAnsCh-F
back propagation artificial neural network Levenberg Marquart arithmetic LD 50 over fitting Hansch Fujita
