Application of Functional‐Link Net in QSAR. 1. QSAR for Activity Data Given by Continuous Variate
Qian Liu, Shuichi Hirono, IKUO MORIGUCHI · Quantitative Structure-Activity Relationships · 1992
Abstract We attempted to apply a new pattern recognition method called “functional‐link net” (Klassen and Pao, 1988) to QSAR. In contrast to the linear weighting produced by the generalized delta rule net often used in neural nets, the functional‐link net acts on a pattern element (a structural parameter in QSAR) or on the entire pattern itself to generate a set of nonlinear functions. In usual QSAR studies, linear forms of parameters are generally used. But, in many cases, parameters might contribute semilinearly to activity. Such a semilinear contribution of parameters can be examined by semilinearly transforming the parameters into a new parameter vector using a procedure with the architecture of the functional‐link net. The new method presented here was devised for analysis of QSAR from activity data given by a continuous variate. The application of this method to QSAR of several data sets of carboquon analogues with antileukemic activity gave better results than those given by multiple regression analysis of the same data sets. The comparison of the results with those given by the generalized delta rule net also showed that FUNCLINK was superior in the predictive ability.