ARTIFICIAL NEURAL NETWORKS APPLIED TO THE QUANTITATIVE STRUCTURE-ACTIVITY RELATIONSHIP STUDY OF DIHYDROPTERIDINE REDUCTASE INHIBITORS

Three-layer artificial neural network models with back-propagation of error have been used to investigate the quantitative structure-activity relationship of dihydropteridine reductase inhibitors, which were derived from the structure of 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine. The network's architecture and parameters were optimized to give maximum performance and an empirical rule for dynamically adjusting the network's learning rate was put forward. The network gave satisfactory results by using f(x) = 1/(1 + e(-x)) as the input-output transformation function of network nodes and setting the number of hidden nodes to six. The results compared favourably with those obtained by stepwise multidimensional linear regression analysis. The concept of partial correlation index, which is similar to the concept of multiple correlation coefficient in multivariate statistical treatment of linear data, was introduced as a measure of the degree of influence of individual input variables on the network's output. It was found that this index worked well for representing the relative contribution of individual structural parameters to the biological activity.