Neural networks applied to pharmaceutical problems. V. Obtaining the correlation indices between drug activity and structural parameters using a neural network.

Abstract

A linear operation was introduced to the neural network. We also studied a method that gives correlation coefficients between input parameters and outputs. It was difficult to obtain these coefficients. Instead, we proposed a new index, the divided difference of output intensity with respect to the input parameter (δO/δx). The divided differences were obtained using an unsaturated sigmoid function as an output function of neurons, where a new learning method with the concept of "neuron fatigue" is introduced. The new techniques were applied to quantitative structure-activity relationships (QSAR) studies in carboquinones and benzodiazepines to compare the results with those by multiregression analysis. It was found that the divided differences work well and can produce the equivalent partial differential coefficients that are obtained by multiregression analysis.