论文部分内容阅读
提出一种线性与非线性相结合的组合式神经网络，用以研究化合物的定量构效关系。网络线性部分的参数和局部输出值可以解释分子结构片段对化合物活性效应的贡献，而其非线性部分则可进一步提高网络模型的拟合与预报精度。还提出了一种快速的类随机搜索训练方法，并实际应用于含硫苯衍生物构效关系的建模中，取得了良好的效果。
A combined neural network combining linearity and nonlinearity was proposed to study the quantitative structure-activity relationship of the compounds. The linear part of the network parameters and local output value can explain the contribution of the molecular structure fragment to the compound activity effect, and its non-linear part can further improve the fitting and forecasting accuracy of the network model. A fast class training method based on stochastic search is also proposed. It is also applied in the modeling of the structure-activity relationship of sulfur-containing benzene derivatives and has achieved good results.