Key points are not available for this paper at this time.
Metal organic frameworks (MOFs) have attracted attention for application of drug delivery because of their ordered porous properties. Optimizing both drug loading capacity and biocompatibility remains a complex challenge for MOFs because these performance indicators depend on nonlinear interactions among structural, compositional, and physicochemical features. In this study, an explainable ensemble learning framework was developed to predict Drug Loading Capacity (g/g) and Cell Viability (%) of drug-loaded MOFs using curated structural descriptors and experimentally derived data. A structured preprocessing pipeline—including robust scaling, multicollinearity control, target-guided encoding, and exploratory dimensionality analysis—was implemented prior to model training. Three gradient boosting algorithms, Gradient Boosting Trees (GBT), Extreme Gradient Boosting (XGBoost), and Histogram-Based Gradient Boosting (HGB), were systematically evaluated. Among them, HGB demonstrated superior predictive performance, achieving test R 2 values of 0.9924 for drug loading capacity and 0.9987 for cell viability, with minimal generalization gap between training and testing datasets. An ablation study confirmed that model performance arises from the synergistic contribution of preprocessing strategies and boosting architecture. Furthermore, SHAP and LIME analyses provided both global and local interpretability, revealing chemically meaningful feature contributions and enhancing model transparency. The results demonstrate that explainable gradient boosting models can reliably capture complex structure–property relationships in MOF systems, offering a robust and interpretable computational tool to accelerate data-driven optimization of drug delivery platforms while reducing reliance on extensive experimental screening.
Alshehri et al. (Thu,) studied this question.