A strategy for developing accurate quantitative structure-activity relationship models enabling predictions of biological properties, when suitable knowledge concerning both ligands and biological target is available, was tested on a data set where molecules are characterized by high structural diversity. Such a strategy was applied to human ether-a-go-go-related gene K+ channel inhibition and consists of a combination of ligand- and structure-based approaches, which can be carried out whenever the three-dimensional structure of the target macromolecule is known or may be modeled with good accuracy. Molecular conformations of ligands were obtained by means of molecular docking, performed in a previously built theoretical model of the channel pore, so that descriptors depending upon the three-dimensional molecular structure were properly computed. A modification of the directed sphere-exclusion algorithm was developed and exploited to properly splitting the whole dataset into Training/Test set pairs. Molecular descriptors, computed by means of the codessa program, were used for the search of reliable quantitative structure-activity relationship models that were subsequently identified through a rigorous validation analysis. Finally, pIC50 values of a prediction set, external to the initial dataset, were predicted and the results confirmed the high predictive power of the model within a quite wide chemical space.