Using Neural Networks and Ensemble Techniques based on Decision Trees for Skin Permeability Prediction

Development of an accurate skin permeability model is becoming increasingly important as skin has been more utilized in recent development of drug delivery methods. This paper presents results of development of Artificial Neural Network (ANN) for prediction of skin permeability. The performance of developed ANN was compared to three regression algorithms used in this paper. The prediction of skin permeability is based on three input parameters: molecular weight, partition coefficient - log(P), and melting temperature for each drug. The dataset of 400 samples was used for prediction of skin permeability. Out of that number, 75% was used for training of ANN, and testing of developed ANN was performed on 100 samples from the dataset. During testing, system correctly predicted 76.7%. This dataset was also used as input to three ensemble techniques based on decision trees: REP Tree, Bagging, Random SubSpace Developed. It was shown that Bagging algorithm outperformed developed ANN with 81% while Random Subspace performed at 73.3%. System can be used in laboratory conditions and can be used in the future for drug discovery.