We describe the application of genetic programming, an evolutionary computing method, to predicting whether small molecules will block the HERG cardiac potassium channel. Models based on a molecular fragment-based descriptor set achieve an accuracy of 85–90% in predicting whether the IC50 of a ‘blind’ set of compounds is <1 μM. Analysis of the models provides a ‘meta-SAR’, which predicts a pharmacophore of two hydrophobic features, one preferably aromatic and one preferably nitrogen-containing, with a protonatable nitrogen asymmetrically situated between them. Our experience of the approach suggests that it is robust, and requires limited scientist input to generate valuable predictive results and structural understanding of the target.
