Particle swarm optimization and neural network application for QSAR

Summary form only given. A successful approach to building QSAR models was proposed by other researchers. It uses binary particle swarm optimization (BPSO) for feature selection in the first stage, and a back propagation neural network in the second stage to generate a QSAR model based on the features selected in the first stage. We start by reestablishing the results of this approach on an extended number of data sets. A new method is then proposed that addresses the limitation of back propagation. This approach uses particle swarm optimization (PSO) in the second stage for training and bootstrap aggregation (bagging) in order to overcome the instability of PSO. The proposed approach yields robust QSAR models, while reducing the variability due to the choice of the back propagation parameters.