The application of stochastic machine learning methods in the prediction of skin penetration

Improving predictions of skin permeability is a significant problem for which mathematical solutions have been sought for around twenty years. However, the current approaches are limited by the nature of the models chosen and the nature of the dataset. This is an important problem, particularly with the increased use of transdermal and topical drug delivery systems. In this work, we apply K-nearest-neighbour regression, single layer networks, mixture of experts and Gaussian processes to predict the skin permeability coefficient of penetrants. A considerable improvement, both statistically and in terms of the accuracy of predictions, over the current quantitative structure-permeability relationships (QSPRs) was found. Gaussian processes provided the most accurate predictions, when compared to experimentally generated results. It was also shown that using five molecular descriptors - molecular weight, solubility parameter, lipophilicity, the number of hydrogen bonding acceptor and donor groups - can produce better predictions than when using only lipophilicity and the molecular weight, which is an approach commonly found with QSPRs. The Gaussian process regression with five compound features was shown to give the best performance in this work. Therefore, Gaussian processes would appear to provide a viable alternative to the development of predictive models for skin absorption and underpin more realistically mechanistic understandings of the physical process of the percutaneous absorption of exogenous chemicals. (C) 2010 Elsevier B.V. All rights reserved.