Hybrid computational model for analysis of drug release from biomaterial matrix: Machine learning aided material design
Haoyu Wang, Zenan Tian, Long Wang, Haifan Wang, Yuxing Zhang, Wenrui Ban · Case Studies in Thermal Engineering · 2025
Prediction of drug release from novel drug delivery systems is of great importance to enhance the efficiency of therapeutics. This study evaluates the predictive capabilities of three machine learning models, Gradient Boosting Decision Trees (GBDT), Deep Neural Networks (DNN), and Neural Oblivious Decision Ensembles (NODE) for modeling complex nonlinear relationships in a dataset containing drug concentration. The data is extracted from mass transfer simulation of drug release from a biomaterial matrix, and the main output is spatial concentration distribution of the drug inside the matrix. Key input features (r and z) ranged from 0.001–0.003 m and 0–0.006 m, respectively, predicting the output concentration (C) within 0.00038–0.000831 mol/m 3 . The models were optimized using the Stochastic Fractal Search (SFS) algorithm, which systematically adjusted hyperparameters to maximize cross-validation R 2 scores. NODE outperformed other models, achieving R 2 scores of 0.99881 (train), 0.99776 ± 0.00003 (validation), and 0.99829 (test), with minimal error metrics: RMSE of 0.00000344 (train) and 0.00000421 (test). In contrast, GBDT and DNN achieved test R 2 scores of 0.97117 and 0.97165, respectively. NODE's predictions were highly consistent across folds, demonstrating robustness and generalizability. SHAP analysis indicated the dominant influence of input feature r, while z contributed negatively to the output. The NODE model is the most appropriate method for handling complex interactions in tabular data thanks to its unique combination of interpretability, scalability, and high predictive performance.