Selective Descriptor Pruning for QSAR/QSPR Studies Using Artificial Neural Networks

Abstract

Selection of optimal descriptors in quantitative structure-activity-property relationship (QSAR/QSPR) studies has been a perennial problem. Artificial Neural Networks (ANNs) have been used widely in QSAR/QSPR studies but less widely in descriptor selection. The current study used ANNs to select an optimal set of descriptors using large numbers of input variables. The effects of clean, noisy, and random input descriptors with linear, nonlinear, and periodic data on synthetic and real data QSAR/QSPR sets were examined. The optimal set of descriptors could be determined using a signal-to-noise ratio method. The optimal values for the rho parameter, which relates sample size to network architecture, were found to vary with the type of data. ANNs were able to detect meaningful descriptors in the presence of large numbers of random false descriptors.