Rational Design of Safer Inorganic Nanoparticles via Mechanistic Modeling-Informed Machine Learning.

Abstract

The safety of inorganic nanoparticles (NPs) remains a critical challenge for their clinical translation. To address this, we developed a machine learning (ML) framework that predicts NP toxicity both <i>in vitro</i> and <i>in vivo</i>, leveraging physicochemical properties and experimental conditions. A curated <i>in vitro</i> cytotoxicity dataset was used to train and validate binary classification models, with top-performing models undergoing explainability analysis to identify key determinants of toxicity and establish structure-toxicity relationships. External testing with diverse inorganic NPs validated the predictive accuracy of the framework for <i>in vitro</i> settings. To enable organ-specific toxicity predictions <i>in vivo</i>, we integrated a physiologically based pharmacokinetic (PBPK) model into the ML pipeline to quantify NP exposure across organs. Retraining the ML models with PBPK-derived exposure metrics yielded robust predictions of organ-specific nanotoxicity, further validating the framework. This PBPK-informed ML approach can thus serve as a potential alternative approach to streamline NP safety assessment, enabling the rational design of safer NPs and expediting their clinical translation.