用BP神经网络预测化合物的经皮渗透速率

Prediction of Percutaneous Permeation of Compounds with Back-Propagation Neural Network

以文献报道的40个化合物的相对分子质量(Mr)、油水分配系数的对数值(1gKow)、供氢数(Hd)和受氢数(Ha)为输入变量,化合物经皮渗透速率对数值(lgJmax)为输出变量,建立了BP神经网络,并用其预测了另8个化合物的lgJmax。结果表明,用预测值对实测值进行线性回归,所得相关系数的平方为0.997。与Potts-Guy模型和Lien-Gao模型相比,BP神经网络得到的预测值与实测值更接近。

The back-propagation （BP） neural network was established using molecular weight （Mr）, logarithm of oil/water partition coefficient （1gKow）, hydrogen-bond donor （Hd） and hydrogen-bond acceptor （Ha） of forty compounds reported in literature as the input variables and logarithm of percutaneous permeation （1gJmax） as the output variable. The 1gJmax of the other eight compounds were predicted by this BP neural network. The results showed that the predicted value of 1gJmax was correlated with the observed value with the squared value of correlation coefficient of 0.997. The 1gJmax of eight compounds predicted by BP neural network were more close to the observed values by comparison with Potts-Guy and Lien-Gao models.