Interpretable Machine Learning for Identifying the Key Factors Controlling Nanocarrier Entrapment Efficiency

Accurately predicting entrapment efficiency (EE) in nanocarrier formulations remains a central challenge in formulation science due to the nonlinear interplay between molecular descriptors, compositional ratios, and process parameters. In this study, a curated dataset of 127 nanocarrier formulations was modeled using seven machine learning regressors including Decision Tree, AdaBoost, Random Forest, K-Nearest Neighbors (KNN), Support Vector Regression (SVR), Linear Regression, and an ensemble learning strategy to identify the most reliable predictive approach for EE estimation. Model performance was evaluated using R², mean squared error (MSE), and average absolute relative error (AARE%) across training, test, and total datasets. The results demonstrate that AdaBoost, Decision Tree, and Ensemble Learning achieved the strongest predictive performance, with test-set R² values of 0.924, 0.900, and 0.914, respectively, and correspondingly low MSE values. These models consistently outperformed Random Forest and KNN, which showed moderate predictive capability (test R² = 0.811 and 0.765). SVR and Linear Regression exhibited limited suitability for this nonlinear formulation space, reflected by substantially lower R² values (0.508 and 0.054) and higher AARE%. The superior performance of tree-based and boosting models highlights their ability to capture complex, multivariate interactions governing EE. The findings confirm that boosting-based and ensemble approaches provide the most robust and accurate predictive framework for EE estimation in heterogeneous nanocarrier systems. These results support the integration of data-driven modeling into formulation design workflows, enabling more efficient screening, reduced experimental burden, and improved mechanistic understanding of the physicochemical determinants of entrapment efficiency.

Hali tarjima qilinmagan