(191z) Quantitative Structure-Property Relationship Model for Prediction of Octanol-Water Partition Coefficient

The octanol-water partition coefficient (K_ow) is a thermophysical property describing the hydrophobicity of a molecule as expressed by the ratio of the organic and aqueous solubilities. This property has been found useful in inferring transport mechanisms and molecular distribution in environmental areas and bio-applications. For example, K_ow is a significant property in describing the permeation of a compound through the skin in transdermal drug delivery.

In this study, we present a new quantitative structure-property relationship (QSPR) model for predicting K_ow utilizing a wide range of molecular species data from the PHYSPROP physical property database. Over 800 descriptors were generated for the 11,308 molecules in the database. A wrapper approach, which involves differential evolution combined with neural networks, was implemented to find the optimum set of 50 network descriptors and the network architecture, simultaneously. A 50-33-35-1 neural network architecture was found to result in the least root-mean square error (RMSE) for the training set data. An ensemble network resulted in RMSE values of 0.28 and 0.38 for the training and validation sets, respectively. The ensemble performed reasonably well on an external dataset when compared with other popular K_ow models in the literature. The partial derivatives method was employed to interpret the resulting model.