An integrated prediction model using machine learning and dynamics simulation for the selection of transdermal penetration enhancers.
TL;DR
A rapid screening platform for PEs in transdermal formulations is established using machine learning (ML) approaches based on a self-constructed database and the permeation process of "drug-PE" systems via molecular dynamics simulation (MD) is explored.
Abstract
Penetration enhancers (PEs) are widely used in transdermal drug delivery systems (TDDS) due to their cost-effectiveness and excellent biocompatibility. However, the rapid selection of PEs for TDDS is hindered by two critical challenges: first, in vitro permeation tests (IVPT) are costly and time consumption; second, the datasets for drug-PE matching are relatively limited. To address these issues, this study established a rapid screening platform for PEs in transdermal formulations using machine learning (ML) approaches based on a self-constructed database and further explored the permeation process of "drug-PE" systems via molecular dynamics simulation (MD). Systematic evaluation of five ML algorithms showed that the Light Gradient Boosting Machine (LightGBM) model achieved the optimal predictive performance using 5-fold cross-validation based on random selection and unseen drug-excipient combination splitting, with R2 values of 0.78 ± 0.03 (random split), 0.73 ± 0.01 (unseen PE split), and 0.72 ± 0.06 (unseen drug-PE combination split) on the test sets, respectively. In contrast, performance decreased substantially for the unseen-drug split (R2 = -0.40), indicating that the model performs well primarily for interpolation within the chemical space represented in the dataset but does not reliably extrapolate to entirely unseen drugs. Feature importance analysis revealed a dose-dependent relationship between PE concentration and transdermal efficiency, while the log S of drugs exerted a non-singular effect on transdermal flux. Moreover, transdermal database analysis indicated that drugs with log P values of 2-3, when paired with PEs possessing log P values of 4-6, tended to achieve higher transdermal permeation efficiency. The accuracy of the prediction model was verified by IVPT using two model drugs, berberine and ferulic acid. Furthermore, MD simulations were employed to elucidate the penetration mechanism of the model drugs in "drug-PE" combinations at the atomic level. Finally, a rapid screening platform (https://permapp-online.streamlit.app) was developed, which enabled transdermal flux prediction through the input of molecular descriptors. This platform can provide valuable support for the future design of transdermal formulations.