An Optimal Self-Pruning Neural Network and Nonlinear Descriptor Selection in QSAR

Feature selection is an important but still poorly solved problem in QSAR modeling. We employ a Bayesian regularized neural network with a sparse Laplacian prior as an efficient method for supervised feature selection, and robust parsimonious nonlinear QSAR modeling. The method simultaneously selects the most relevant descriptors for model, and automatically prunes the neural network to have the architecture with optimum prediction ability. We illustrate the advantages of the method using a suite of diverse data sets, and compare the results obtained by the new method against those obtained by alternative contemporary methods.