Negative examples, which are required for most machine learning methods to infer new predictions, are rarely directly recorded in several real world databases for classification problems. A variety of heuristics for the choice of negative examples have been proposed, ranging from simply under-sampling non positive instances, to the analysis of class taxonomy structures. Here we propose an efficient strategy for selecting negative examples designed for Hopfield networks which exploits the clustering properties of positive instances. The method has been validated on the prediction of protein functions of a model organism.