Predicting nanoparticle tumor delivery efficiency using explainable machine learning for cancer nanomedicine design

Nanoparticle-based drug delivery systems have been proposed as a promising strategy for improving the efficacy and specificity of cancer therapeutics. However, predicting tumor delivery efficiency remains challenging because of complex interactions among nanoparticle physicochemical properties, biological environments, and experimental conditions. This study presents an explainable machine-learning framework for predicting nanoparticle tumor delivery efficiency using the updated Nano-Tumor Database. Following data preprocessing and feature engineering, 2,319 records were used to develop and evaluate Random Forest, Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), weighted ensemble, and stacking ensemble models. To reduce information leakage from repeated nanoparticle measurements, model development and evaluation were revised using nanoparticle-identity-aware grouping based on NM.ID, with categorical preprocessing fitted exclusively on training data. In the final untouched NM.ID-grouped hold-out set, the stacking ensemble achieved the lowest RMSE (3.8107) and highest $R^2$ (0.0901), while the weighted ensemble achieved the lowest MAE (1.8320); however, differences among the leading models were small. Five-fold NM.ID-grouped cross-validation yielded mean $R^2$ values of 0.2280 for XGBoost, 0.2169 for Random Forest, and 0.1980 for LightGBM, demonstrating that generalization to previously unseen nanoparticle identities is substantially more challenging than suggested by row-wise validation. Repeated group-aware splits further showed variability in model ranking, while stricter reference-grouped validation indicated limited cross-study transportability. SHapley Additive exPlanations (SHAP) were used to examine model interpretability, with Animal, Tumor Cell, Core material, Dose, and Zeta Potential showing the largest global contributions to predictions from the final tuned XGBoost model. Dose showed no meaningful marginal correlation with tumor %ID, although group-aware ablation indicated modest multivariable predictive contribution. Performance generally deteriorated for core-material categories absent from training, indicating that the framework should primarily be interpreted as an explainable screening and hypothesis-generation tool within the represented chemical and experimental domain rather than as a validated predictor for completely novel nanoparticle chemistries. These findings highlight both the potential and current generalization limitations of data-driven approaches for supporting nanoparticle design in cancer nanomedicine.