Artificial neural networks (ANNs) are often used "in silico" drug design for correlation, classification, and prediction of the activity of bioactive compounds and drug. In this chapter, we approach the implementation of Bayesian-regularized artificial neural networks (BRANNs) combined with genetic algorithm (GA) feature selection, so-called Bayesian-regularized genetic neural networks (BRGNNs), in quantitative structure-activity relationships (QSAR) analysis. BRGNN was applied to the computer-aided design of drug candidates for a variety of diseases, that is, cancer, AIDS, fungal infections, cardiac disease, etc. The neural networks were successfully trained to calculate the biological activities of a wide spectrum of drug candidates using different levels of representation of the chemical information. 2D and 3D structural descriptors were more frequently used, but quantum chemical descriptors also yielded good neural network models. In general, GA feature selection improves previous approaches by being more accurate and robust. In addition, we describe how feature selection over large pools of molecular descriptors provided valuable structural insights into ligand-target interactions.