摘要
药物定量构效关系一直是药物现代化研究中的关键问题。现已发现具有一定特征或一定作用机制的同源活性化合物,其主要结构因素与活性强度间存在着一定的关系。通过运用计算机建模仿真、神经元计算等方法,可以搜索化学组成与药效相随变动规律,以及预测未知化合物的生物活性。本文将内部递归神经网络应用到药物定量构效关系建模中。为了提高其网络的性能,本文提出了改进型内部递归神经网络,具体策略包括:运用动量法和自适应学习率改进网络的学习能力,采用提早结束法改善网络的泛化能力,并成功地将其应用于除草剂的定量构效关系建模中。
QSAR is always the key issue in the modemization of medicine. It is found that there is a relationship between structure and active intensity of the homologous and active compound, which has some definite characteristics and active mechanisms. With use of computer modeling simulation and neural computing, we can search the rule between chemical composition and drug effect, and predict the bioactivity of unknown component. The intemal recurrent neural network is employed to predict the activity of chemical medicine in this paper. In order to enhance the performance of this network, a modified recurrent neural network （MRNN） is proposed. The strategy of MRNN is that momentum and self-adapting learning rate are applied in training RNN to improve the leaming ability, and early stopping training is applied to promote generalization ability. The MRNN is successfully applied to quantitative strueture- activity relationship study of herbicide.
出处
《计算机与应用化学》
CAS
CSCD
北大核心
2007年第2期196-200,共5页
Computers and Applied Chemistry
关键词
化学药品
内部递归神经网络
动态学习率
泛化能力
chemical medicine, internal recurrent neural network, self-adapting leaming, generalization ability
