主成分分析的神经网络及其在木聚糖酶最适pH建模中的应用

A principal component analysis based output integrated BP neural network for xylanase optimum pH modeling

木聚糖酶结构与功能、性质的关系错综复杂,传统的回归分析往往不能满足要求。本文采用主成分分析法对样本数据集进行预处理,将得到的新样本数据集输入神经网络,籍助于均匀设计(UD),构建了木聚糖酶氨基酸组成和最适pH的模型。当学习速率为0.08、动态参数为0.7、Sigmoid参数为0.92,隐含层结点数为9时,模型的拟合残差为0.001 09,对pH值拟合的平均绝对百分比误差为3.29%,同时具有良好的预测效果,预测的平均绝对误差为0.59个pH单位。比文献报道的用逐步回归方法更好。

The structure-fun,：tion and structure-activity relationship of xylanases was complicated, and the conventional regression methods usually can not yield a satisfied solution to it. The principal component analysis was applied to the data processing in training sets, the new principal components were then used as input parameters of BP neural networks. A prediction model for optimum pH of xylanase in G/11 family was established based on uniform design. When the learning rate （η） , momentum parameter, Sigmoid parameter and the neuron numbers of the hidden layer were 0.08, 0.7, 0.92 and 9, the calculated pHs fitted the reported optimum pHs of xylanase very well and the MAPEs （ Mean Absolute Percent Error） was 3.29% . At the same time, the predicted pHs fitted the reported optimum pHs well and the MAE （ Mean Absolute Error） was 0.59 pH unit. It was superior in fittings and predictions compared to the reported model based on stepwise regression.