QSAR中ANN用于研究变量选择方法的回顾和比较

Review and comparision of methods to study the contribution of variables in artificial neural network models for QSAR study

摘要:

在定量构效关系(QSAR)建模中,人工神经网络(ANN)模型预测能力较好,但并不能提供化合物结构变量对活性影响的更多信息,因而被认为是个黑箱模型.以35种硝基化合物对黑呆头鱼96 h的生物毒性为例,首次回顾和比较6种用于研究网络输入变量对输出相对贡献大小的方法.结果表明:ANN中引入变量选择的方法,大大增强了QSAR模型的解释能力,其中偏微分方法能得出最为全面准确的结果,其次为轮廓图方法.扰动法和权重法对输入参数能实现较好的分类,但过于简化且方法不稳定;而传统的逐步回归法结果最差.

Abstract:

Although Artificial Neural Network(ANN) shows superior predictive power in the study of quantitative structure-activity relationship(QSAR),it has been labeled as a "black box" because it provides little explanatory insight into the relative importance of the independent variables.In this paper,as an example of toxicity of 35 nitro-aromatics on fathead minnow,six methods which could give the relative contribution and/or the contribution profile of input factors were reviewed and compared.The Partial Derivative method was found to be the most useful as it gave the most complete results,followed by the Profile method that gave the contribution profile of input variables.The Perturb method allowed a good classification of input parameters as well as the Weights method that had been simplified,but these two methods lacked stability.Finally,the classical stepwise methods gave the poorest results.