Over the past decade, the pharmaceutical industry has begun to address an addition to ADME/Tox profiling--the ability of a compound to bind to and inhibit the human ether-a-go-go-related gene (hERG)-encoded cardiac potassium channel. With the compilation of a large and diverse set of compounds measured in a single, consistent hERG channel inhibition assay, we recognized a unique opportunity to attempt to construct predictive QSAR models. Early efforts with classification models built from this training set were very encouraging. Here, we report a systematic evaluation of regression models based on neural network ensembles in conjunction with a variety of structure representations and feature selection algorithms. The combination of these modeling techniques (neural networks to capture non-linear relationships in the data, feature selection to prevent over-fitting, and aggregation to minimize model instability) was found to produce models with very good internal cross-validation statistics and good predictivity on external data.