Adaptive neuro-fuzzy inference system: an instant and architecture-free predictor for improved QSAR studies

The application of an adaptive neuro-fuzzy inference system (ANFIS) has been developed for obtaining sufficient quantitative structure-activity relationships (QSAR) with high accuracy. To this end, a data set of 68 pyrimidines derivatives as DHFR inhibitors, described first in the excellent independent studies of Hansch et al. (J. Med. Chem. 1982, 25, 777-784 and J. Med. Chem. 1991, 34, 46-54) and later by So and Richards (J. Med. Chem. 1992, 35, 3201-3207), was examined. The ANFIS system, first time applied in the literature to QSAR studies, was trained using a hybrid algorithm consisting of back-propagation and least-squares estimation while the optimum number and shape of membership functions were obtained through the subtractive clustering algorithm. Prior to the development and evaluation of the ANFIS system, geometry optimization of the examined compounds was performed, deriving a series of diverse descriptors from which the best subset was selected by using a hybrid genetic algorithm system. The predictive abilities of the resulting models compared to those produced from classical multivariate regression such as linear and nonlinear (quadratic) partial least squares regression (PLS and QPLS, respectively). The ANFIS method outperformed both the PLS models as well as the published results, leading to substantial gain in both the prediction ability and the computation speed (almost instant training).