Classification of Potent and Weak Penetration Enhancers Using Multiple Feature Selection Methods and Machine Learning Models

Purpose Chemical penetration enhancers (CPEs) are important in transdermal drug delivery (TDDD) formulations because they assist drugs in moving across the stratum corneum. Hydrocortisone (0.1% hydrocortisone, propylene glycol), oestradiol (0.045 mg estradiol/0.015 mg levonorgestrel, propylene glycol), and testosterone (2% testosterone, propylene glycol) are some examples of marketing TDDD formulations. As the transdermal route for drug administration becomes a safer and more appealing alternative to hypodermic needles, the search for new CPEs and their development becomes more important. Thus, the current work was directed toward the rapid identification of potent CPEs through the development of robust machine learning (ML) classification models. Methods Two large penetration enhancer (PE) data sets reported to date such as hydrocortisone (139 PEs) and theophylline (101 PEs) were used to build classification models. In the present investigation, a combination of feature selection methods, i.e., Boruta and Recursive Feature Elimination (RFE), and machine learning (ML) algorithms such as support vector machine (SVM), random forest (RF), and artificial neural network (ANN) were employed to classify the potent and weak penetration enhancers of hydrocortisone and theophylline. The tenfold cross-validation and Y-randomization methods were used to evaluate the prediction performance of the developed models. Results Significant classification models were built for both data sets when the RFE method and RF algorithm were used. RF classifiers outperformed hydrocortisone and theophylline data sets with test set accuracy and Matthew's correlation coefficient (MCC) greater than 0.78. Simultaneously, four important features required for the accurate classification of potent and weak PEs were identified, i.e., nHCsatu, minHCsatu, AATS4p, and GATS4e. Conclusion Our approach produced robust ML classification models that can be applied to prioritize PEs from large databases. Utilization of these ML models in virtual screening experiments could save time and effort in the identification of potential PEs.