Development of Quantitative Structure-Activity Relationships and Classification Models for Anticonvulsant Activity of Hydantoin Analogues
Development of Quantitative Structure-Activity Relationships and Classification Models for Anticonvulsant Activity of Hydantoin Analogues
复制标题
乙内酰脲类似物抗惊厥活性的定量构效关系和分类模型的开发
DOI：
10.1021/ci025639w
复制
发表时间：
2003
期刊：
影响因子：
--
通讯作者：
D. Weaver
中科院分区：
文献类型：
--
作者：
J. Sutherland;D. Weaver
对在小鼠和大鼠身上测得的大量具有抗惊厥活性的海因衍生物进行了分类和QSAR分析。分类集包括287个具有最大电休克(MES)活性的海因，以定性形式表示。94个具有MES ED(50)值的海因的子集用于QSAR分析。数值描述符用来编码分子的拓扑、几何/结构、电子和热力学性质。使用不同化合物在主成分空间中的表示来选择不同化合物的训练集和测试集进行分析。在这个过程中采用了基于单元和距离度量的选择方法。对于QSAR，使用遗传算法(GA)来选择5-9个描述符的子集，以最小化训练集上的均方根误差。最具预测性的模型在基于单元和距离度量的测试集上的均方根误差分别为0.86(r(2)=0.64)和0.73(r(2)=0.75)ln(1/ED(50))个单位，并且在选定的描述符中表现出收敛。分类模型使用递归划分(RP)和带遗传算法的样条线拟合(SFGA)，这是我们已经实现的一种新方法。最具预测性的RP和SFGA模型在测试集上的分类率分别为75%和80%；这两种方法产生的模型具有相似的区分特征。对于定量构效关系和分类，共识方案提供了更高的预测精度。
Classification and QSAR analysis was performed on a large set of hydantoin derivatives with measured anticonvulsant activity in mice and rats. The classification set comprised 287 hydantoins having maximal electroshock (MES) activity expressed in qualitative form. A subset of 94 hydantoins with MES ED(50) values was used for QSAR analysis. Numerical descriptors were generated to encode topological, geometric/structural, electronic, and thermodynamic properties of molecules. Analyses were performed with training and test sets of diverse compounds selected using their representation in a principal component space. Cell- and distance metric-based selection methods were employed in this process. For QSAR, a genetic algorithm (GA) was used for selecting subsets of 5-9 descriptors that minimize the rms error on the training sets. The most predictive models have rms errors of 0.86 (r(2) = 0.64) and 0.73 (r(2) = 0.75) ln(1/ED(50)) units on the cell- and distance metric-derived test sets, respectively, and showed convergence in the selected descriptors. Classification models were developed using recursive partitioning (RP) and spline-fitting with a GA (SFGA), a novel method we have implemented. The most predictive RP and SFGA models have classification rates of 75% and 80% on the test sets; both methods produced models with similar discriminating features. For QSAR and classification, consensus schemes gave improved predictive accuracy.