贝叶斯规整化神经网络模型预测吲哚烷胺类化合物对5-HT_(1B/1D)受体亲和力  

The Bayesian Regularized Neural Network Prediction Model for Indolealkylamines Affinity to 5-HT_(1B/1D) Receptors

作　　者：温秋玲[1] 杨博[1] 戴康[1] 

机构地区：[1]中南民族大学药学院,武汉430074

出　　处：《医药导报》2010年第5期555-558,共4页Herald of Medicine

基　　金：国家自然科学基金资助项目(基金编号:30471432)

摘　　要：目的利用贝叶斯规整化神经网络模型研究45种吲哚烷胺对5羟-色胺(5-HT)1B/1D受体激动活性的定量构效关系。方法选取115种与结构相关的拓扑参数、几何参数和疏水性参数等,通过主成分分析法进行参数减元,建立基于10种主成分变量的活性预测贝叶斯规整化神经网络模型,并利用去一法(LEAVE-ONE-OUT)对模型进行交叉验证。结果应用残差绝对值的平均值(MAE)进行筛选,得到隐含层神经元数目为10的模型为最佳模型。在该模型下,吲哚烷胺对5-HT1B受体和5-HT1D受体亲和力的实验值和预测值一元相关系数平方(R2)分别为0.990 5和0.988 7。结论模型显示吲哚烷胺5-HT受体激动作用与其结构有密切关系。贝叶斯规整化神经网络结合主成分分析方法有良好的预测能力,且稳定可靠,有望在5-HT1B/1D受体激动药新药设计中得到广泛应用。Objective A quantitative structure-activity relationship （QSAR） model was built on the base of 45 indolealkylamines and by the method of Bayesian regularized neural network （BRNN）. This model was used to predict activities （anti-migraine） of 5-HT1B/1D receptor agonists. MethodsOne hundred and fifteen QSAR indices were used to elucidate the structural characters of indolealkylamines, which included topological,geometrical and hydrophobicity indices and so on. 115 variables were compressed by principle component analysis（PCA）, and several PCA variables were chosen as input ones of model. BRNN model was trained and LEAVE-ONE-OUT （LOO） was used as a cross-validation method. ResultsIn the present study, we obtained an optimized model with ten nerve centers in the hidden layer. The correlation coefficience for predicted vs experimental value was 0.990 5 for 5-HT1B and 0.988 7 for 5-HT1D. ConclusionsThe results indicate that the structures of indolealklyamines have good relationship with the affinity to the receptors. They also show that the BRNN model exhibites a good predictive ability and is robustness. We expect that this method may be widely applied in the area of new drug design for anti-migraine.

关 键 词：吲哚烷胺 5羟-色胺1B/1D受体 定量构效关系 贝叶斯规整化神经网络 

分 类 号：R965[医药卫生—药理学] R971[医药卫生—药学]