Andrea T A, Kalayeh H
E. I. du Pont de Nemours, Stine-Haskell Research Center, Newark, Delaware 19711.
J Med Chem. 1991 Sep;34(9):2824-36. doi: 10.1021/jm00113a022.
Back propagation neural networks is a new technology useful for modeling nonlinear functions of several variables. This paper explores their applications in the field of quantitative structure-activity relationships. In particular, their ability to fit biological activity surfaces, predict activity, and determine the "functional forms" of its dependence on physical properties is compared to well-established methods in the field. A dataset of 256 5-phenyl-3,4-diamino-6,6-dimethyldihydrotriazines that inhibit dihydrofolate reductase enzyme is used as a basis for comparison. It is found that neural networks lead to enhanced surface fits and predictions relative to standard regression methods. Moreover, they circumvent the need for ad hoc indicator variables, which account for a significant part of the variance in linear regression models. Additionally, they lead to the elucidation of nonlinear and "cross-products" effects that correspond to trade-offs between physical properties in their effect on biological activity. This is the first demonstration of the latter two findings. On the other hand, due to the complexity of the resulting models, an understanding of the local, but not the global, structure-activity relationships is possible. The latter must await further developments. Furthermore, the longer computational time required to train the networks is somewhat inconveniencing, although not restrictive.
反向传播神经网络是一种可用于对多个变量的非线性函数进行建模的新技术。本文探讨了它们在定量构效关系领域的应用。特别是，将它们拟合生物活性表面、预测活性以及确定其对物理性质依赖性的“函数形式”的能力与该领域中已确立的方法进行了比较。使用一个包含256种抑制二氢叶酸还原酶的5-苯基-3,4-二氨基-6,6-二甲基二氢三嗪的数据集作为比较基础。结果发现，相对于标准回归方法，神经网络能实现更好的表面拟合和预测。此外，它们无需特设指示变量，而在线性回归模型中，特设指示变量占方差的很大一部分。此外，它们还能揭示非线性和“交叉乘积”效应，这些效应对应于物理性质对生物活性影响之间的权衡。这是后两个发现的首次证明。另一方面，由于所得模型的复杂性，只能理解局部而非全局的构效关系。全局构效关系还有待进一步发展。此外，训练网络所需的较长计算时间虽不构成限制，但也有些不便。
