Prediction of human skin permeability using artificial neural network (ANN) modeling

Aim: To develop an artificial neural network (ANN) model for predicting skin permeability (log K-p) of new chemical entities. Methods: A large dataset of 215 experimental data points was compiled from the literature. The dataset was subdivided into 5 subsets and 4 of them were used to train and validate an ANN model. The same 4 datasets were also used to build a multiple linear regression (MLR) model. The remaining dataset was then used to test the 2 models. Abraham descriptors were employed as inputs into the 2 models. Model predictions were compared with the experimental results. In addition, the relationship between log K-p and Abraham descriptors were investigated. Results: The regression results of the MLR model were n=215, determination coefficient (R-2)=0.699, mean square error (MSE)=0.243, and F=493.556. The ANN model gave improved results with n=215, R-2=0.832, MSE=0.136, and F=1050.653. The ANN model suggests that the relationship between log K-p and Abraham descriptors is non-linear. Conclusion: The study suggests that Abraham descriptors may be used to predict skin permeability, and the ANN model gives improved prediction of skin permeability.