Artificial neural network analysis for predicting human percutaneous absorption taking account of vehicle properties.

An in silico method for predicting percutaneous absorption of cosmetic ingredients was developed by using artificial neural network (ANN) analysis to predict the human skin permeability coefficient (log Kp), taking account of the physicochemical properties of the vehicle, and the apparent diffusion coefficient (log D). Molecular weight and octanol-water partition coefficient (log P) of chemicals, and log P of the vehicles, were used as molecular descriptors for predicting log Kp and log D of 359 samples, for which literature values of either or both of log Kp and log D were available. Adaptivity of the ANN model was evaluated in comparison with a multiple linear regression model (MLR) by calculating the root-mean-square (RMS) errors. Accuracy and robustness were confirmed by 10-fold cross-validation. The predictive RMS errors of the ANN model were smaller than those of the MLR model (log Kp; 0.675 vs 0.887, log D; 0.553 vs 0.658), indicating superior performance. The predictive RMS errors for log Kp and log D with the ANN model after 10-fold cross-validation analysis were 0.723 and 0.606, respectively. Moreover, we estimated the cumulative amounts of chemicals permeated into the skin during 24 hr (Q24hr) from the values of log Kp and log D by applying Fick's law of diffusion. Our results suggest that this newly established ANN analysis method, taking account of the property of the vehicle, could contribute to non-animal risk assessment of cosmetic ingredients by providing a tool for calculating Q24hr, which is required for evaluating the margin of safety.