Octanol-water partition coefficient (K-ow) is an important thermodynamic property used to characterize the partitioning of solutes between an aqueous and organic phase and has importance in such areas as pharmacology, pharmacokinetics, pharmacodynamics, chemical production and environmental toxicology. We present a non-linear quantitative structure-property relationship model for determining K-ow values of new molecules in silico. A total of 823 descriptors were generated for 11,308 molecules whose K-ow values are reported in the PhysProp dataset by Syracuse Research. Optimum network architecture and its associated inputs were identified using a wrapper-based feature selection algorithm that combines differential evolution and artificial neural networks. A network architecture of 50-33-35-1 resulted in the least root-mean squared error (RMSE) in the training set. Further, to improve on single-network predictions, a neural network ensemble was developed by combining five networks that have the same architecture and inputs but differ in layer weights. The ensemble predicted the K-ow values with RMSE of 0.28 and 0.38 for the training set and internal validation set. respectively. The ensemble performed reasonably well on an external dataset when compared with other popular K-ow models in the literature. (C) 2012 Elsevier B.V. All rights reserved.