Ramírez-Galicia Guillermo, Garduño-Juárez Ramón, Hemmateenejad Bahram, Deeb Omar, Deciga-Campos Myrna, Moctezuma-Eugenio Juan Carlos
Instituto de Ciencias Físicas, Universidad Nacional Autónoma de México, PO Box 48-3, 62250 Cuernavaca, Morelos, Mexico.
Chem Biol Drug Des. 2007 Jul;70(1):53-64. doi: 10.1111/j.1747-0285.2007.00530.x.
Quantitative structure-activity relationship studies were performed to describe and predict the antinociceptive activity of 31 morphinan derivatives reported by the US Drug Evaluation Committee in 2005 and 2006. From these, three data sets were constructed and several models were calculated following the multiple linear regression and Leave-One-Out Cross-Validation (LOO-CV) tests. In general, these models achieved good descriptive power (approximately 92%) as well as predictive power (approximately 76%), but were unable to predict an external validation set of morphinan derivatives. When artificial neural networks were applied to these models, an improvement of the predictive and external validation values was obtained. It was observed that the results of the NN models are significantly better that those obtained by multiple linear regression. In spite that the problem under investigation can be handled adequately by a linear model, a neural network does bring slight improvements in the predictive power.
开展了定量构效关系研究，以描述和预测美国药物评估委员会在2005年和2006年报告的31种吗啡喃衍生物的镇痛活性。从中构建了三个数据集，并在进行多元线性回归和留一法交叉验证（LOO-CV）测试后计算了几个模型。总体而言，这些模型具有良好的描述能力（约92%）和预测能力（约76%），但无法预测吗啡喃衍生物的外部验证集。当将人工神经网络应用于这些模型时，预测值和外部验证值得到了改善。据观察，神经网络模型的结果明显优于多元线性回归得到的结果。尽管所研究的问题可以通过线性模型得到充分处理，但神经网络确实在预测能力方面带来了轻微的改善。
