Development and comparison of machine learning models for in-vitro drug permeation prediction from microneedle patch

Area of Science:

• Pharmacology and Pharmaceutics
• Biomedical Engineering
• Data Science

Background:

• Machine learning (ML) offers potential for optimizing drug delivery system development.
• Predicting drug release patterns prior to fabrication can reduce costs and experimental time.
• Microneedle patches are an advanced drug delivery system requiring accurate release prediction.

Purpose of the Study:

• To develop and evaluate ML models for predicting in-vitro drug release and permeation from microneedle patches.
• To compare the performance of stacking regressor, artificial neural network (ANN), and voting regressor models.
• To assess the accuracy of predicting drug permeation amount versus drug permeation percentage.

Main Methods:

• Utilized a literature-derived dataset to train and evaluate stacking regressor, ANN, and voting regressor models.
• Compared model performance using R-squared score, RMSE, and MAE.
• Optimized hyperparameters for the best-performing model and performed cross-validation.

Main Results:

• The voting regressor model achieved the best performance for drug permeation prediction (RMSE = 3.24), significantly outperforming stacking regressor (RMSE = 16.54) and ANN (RMSE = 14).
• Predicting permeation amount via predicted percentage (RMSE = 654.94) was more accurate than direct amount prediction (RMSE = 669.69).
• All developed models substantially outperformed a previous model (RMSE = 4447.23).

Conclusions:

• Machine learning, particularly the voting regressor model, can accurately predict drug permeation from microneedle patches.
• The developed models offer a significant improvement over existing methods, facilitating efficient microneedle patch development.
• A web application was developed to enable user-friendly prediction of drug permeation, showcasing ML's practical application in drug delivery systems.