End-to-end prediction of skin permeability from SMILES using a hybrid and explainable GNN framework

Alakuş D. O., ALAKUŞ T. B.

Journal of Drug Delivery Science and Technology, cilt.124, 2026 (SCI-Expanded, Scopus)

Özet

Prediction of skin permeability plays a critical role in transdermal drug design, as well as in toxicology and safety assessments. The classic quantitative structure-permeability relationship (QSPR) models are often based on linear approaches such as Potts and Guy equation and use only a limited number of physiochemical parameters. However, these models cannot adequately represent complex and nonlinear molecular interactions that determine the transport of matter from the skin barrier. In recent years, artificial intelligence-based methods, especially machine learning (ML) and deep learning (DL) models have been successfully applied in the field of skin permeability to overcome these limitations. Yet, these approaches are still largely dependent on hand-selected molecular identifiers. In this context, the Graph Neural Networks (GNNs) recently attract attention by representing molecules directly in graphic form and learning inter-atoms relationships. GNN-based models can directly draw properties from the chemical structure and go beyond classic QSPR approaches. This study proposes a hybrid DL frame, which can predict the skin permeability coefficient directly from molecular structures. The model will integrate the architecture of message passing neural network (MPNN) and graph attention networks (GATs) through graphic representations generated from SMILES representations of molecules. The success of the developed hybrid method was calculated with R2, Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Spearman correlation coefficient and the results were compared with Random Forest (RF), XGBoost (XGB), Feedforward Neural Network (FNN), Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) results. At the end of the study, the proposed hybrid MPNN-GAT model showed approximately 9%-25% error decreased compared to existing models and a performance of 0.86 at R2. As a result, the developed hybrid model offered a powerful alternative to the literature beyond the existing methods due to both high estimation accuracy and explainability features. Additionally, a multimodal base model was also included in the study to ensure fair comparison between models with different data representations.