Koutsoukos A D, Rubinstein L V, Faraggi D, Simon R M, Kalyandrug S, Weinstein J N, Kohn K W, Paull K D
Biometric Research Branch, National Cancer Institute, Bethesda, MD 20892.
Stat Med. 1994;13(5-7):719-30. doi: 10.1002/sim.4780130532.
The National Cancer Institute currently tests approximately 400 compounds per week against a panel of human tumour cell lines in order to identify potential anti-cancer drugs. We describe several approaches, based on these in vitro data, to the problem of identifying the primary biochemical mechanism of action of a compound. Using linear and non-parametric discriminant procedures and cross-validation, we find that accurate identification of the mechanism of action is achieved for approximately 90 per cent of a diverse collection of 141 known compounds, representing six different mechanistic categories. We demonstrate that two-dimensional graphical displays of the compounds in terms of the initial three principal components (of the original data) result in suggestive visual clustering according to mechanism of action. Finally, we compare the classification accuracy of the statistical discrimination procedures with the accuracy obtained from a neural network approach and, for our example, we find that the results obtained from the various approaches are similar.
美国国立癌症研究所目前每周针对一组人类肿瘤细胞系测试约400种化合物，以识别潜在的抗癌药物。我们基于这些体外数据，描述了几种用于识别化合物主要生化作用机制问题的方法。使用线性和非参数判别程序以及交叉验证，我们发现对于代表六种不同作用机制类别的141种已知化合物的多样化集合中的约90%，能够准确识别其作用机制。我们证明，根据（原始数据的）最初三个主成分对化合物进行二维图形展示，会根据作用机制产生具有启发性的视觉聚类。最后，我们将统计判别程序的分类准确性与从神经网络方法获得的准确性进行比较，就我们的示例而言，我们发现从各种方法获得的结果相似。
