Activity Prediction of Hormone-Sensitive Lipase Inhibitors Based on Machine Learning Methods
Activity Prediction of Hormone-Sensitive Lipase Inhibitors Based on Machine Learning Methods
复制标题
基于机器学习方法的激素敏感脂肪酶抑制剂的活性预测
DOI：
10.3866/pku.whxb20100125
复制
发表时间：
2010
影响因子：
10.9
通讯作者：
Lue Wei
中科院分区：
文献类型：
--
作者：
Xue Ying;Lue Wei
激素敏感脂肪酶（HSL）是调节脂肪组织游离脂肪酸（FFA）代谢的关键限速酶。近年来，HSL被发现可用于治疗糖尿病，因此发现新的HSL抑制剂（HSLI）是令人感兴趣的。高度期望用于预测HSLI的方法以促进新型糖尿病治疗剂的设计，因为关于酶敏感性脂肪酶的机制和三维（3D）结构存在有限的知识。我们探索了几种机器学习方法（支持向量机（SVM）、k-最近邻（k-NN）和C4.5决策树（C4.5 DT）），以从一组全面的已知HSLI和非HSLI中预测期望的HSLI。我们的预测系统使用252种化合物（123种HSLI和129种非HSLI）进行了测试，这些化合物的化学结构比其他研究中的化合物更加多样化。使用递归特征消除选择方法来提高预测精度，并选择负责区分HSLI和非HSLI的分子描述符。基于独立验证集的3种机器学习方法对HSLI、nonHSLI和所有结构的预测准确率分别为85.7%~ 90.5%、63.2%~ 68.4%和75.0%~ 80.0%。支持向量机给出了最好的总精度为80.0%的所有结构。这项工作表明，机器学习方法，如支持向量机是有用的，以预测潜在的HSLI未知的化合物之间的集合，并表征与HSLI相关的分子描述符。
Hormone-sensitive lipase (HSL) is known as the key rate-limiting enzyme responsible for regulating free fatty acids (FFAs) metabolism in adipose tissue. Recently,HSL has been found to be useful in the treatment of diabetes so the discovery of new HSL inhibitors (HSLIs) is of interest. Methods for the prediction of HSLIs are highly desired to facilitate the design of novel diabetes therapeutic agents because limited knowledge exists concerning the mechanism and three dimensional (3D) structure of hormone-sensitive lipase. We have explored several machine learning methods (support vector machines (SVM),k-nearest neighbor (k-NN),and C4.5 decision tree (C4.5 DT)) to predict desirable HSLIs from a comprehensive set of known HSLIs and non-HSLIs. Our prediction system was tested using 252 compounds (123 HSLIs and 129 non-HSLIs) and these are significantly more diverse in chemical structure than those in other studies. The recursive feature elimination selection method was used to improve the prediction accuracy and to select the molecular descriptors responsible for distinguishing HSLIs and non-HSLIs. Prediction accuracies were 85.7%-90.5% for HSLIs,63.2%-68.4% for non-HSLIs,and 75.0%-80.0% for all structures based on three kinds of machine learning methods using an independent validation set. SVM gave the best total accuracy of 80.0% for all the structures. This work suggests that machine learning methods such as SVM are useful to predict the potential HSLIs among unknown sets of compounds and to characterize the molecular descriptors associated with HSLIs.