Integrating Physiologically Based Pharmacokinetic Modeling with Machine Learning for the Rational Design of Nanoparticles to Improve Safety in Biomedical Applications

OBJECTIVES: Inorganic nanoparticles (NPs) as drug delivery systems have shown remarkable advancements, exhibiting high degrees of adaptability, which enables precise tailoring of their physicochemical properties, such as size, shape, charge, surface chemistry, and composition [1]. This has led to the development of NPs with enhanced functionalities for drug delivery. However, concerns about safety due to off-target accumulation have negatively influenced the clinical success of inorganic NPs [2]. To this end, we present a machine learning (ML)-based NP toxicity prediction model, informed by a minimal physiologically based pharmacokinetic (PBPK) model, to predict organ-specific toxicity and guide rational design of NPs for enhanced safety.METHODS: Through rigorous meta-analysis, an extensive in vitro toxicity dataset was curated, standardized, and split 70/30 to train and validate 22 binary classification models (linear, clustering, tree, and boosting-based), where 10-fold nested cross-validation optimized hyperparameters and ensured generalizability. The in vitro dataset (N&#3f8752) was characterized by ten attributes, including NP physicochemical properties and experimental conditions. The top five performing models were selected for explainability analysis via SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDPs) to quantify structure-toxicity relationships. To enable prospective in vivo NP toxicity predictions, we integrated a seven-compartment PBPK model into the ML models to provide inputs of organ-specific NP exposure based on NP design characteristics. RESULTS: • Model validation yielded exceptional performance from our top five models: CatBoost (AUC = 0.9443), XGBoost (AUC = 0.9411), Gradient Boosting Classifier (AUC = 0.93104), Random Forest (AUC = 0.9288), and Extra Trees (AUC = 0.9250)• SHAP ranked concentration, composition, size, exposure time, organ, surface coating, and Zeta potential (ζ) as most influential in determining toxicity. • PDP quantified the probability of NP toxicity as a function of NP features, highlighting the mathematical correlation for concentration, size, and exposure time. • Models were trained against in vivo data, with input of NP exposure obtained from a PBPK model, leading to the high predictive performance of toxicity models.CONCLUSION: We developed a PBPK-integrated ML framework to predict the organ-specific toxicity of inorganic NPs. SHAP and PDPs determined concentration, size, composition, coating, and ζ to heavily influence NP toxicity, providing guidelines for rational NP design.Citations: [1] A.A. Yetisgin, S. Cetinel, M. Zuvin, A. Kosar, O. Kutlu, Molecules, 25 (2020) 2193.[2] D. Bobo, K.J. Robinson, J. Islam, K.J. Thurecht, S.R. Corrie, Pharm Res, 33 (2016) 2373-2387.