A Kinetic Machine Learning Model to Simulate PM-Induced Cellular ROS Generation Across the Human Respiratory Tract

The generation of reactive oxygen species (ROS), particularly superoxide radical anion (·O₂^-), is a primary driver of particulate matter (PM) toxicity. However, traditional toxicological assessments are often limited by static endpoint measurements and high experimental doses, failing to capture the dynamic reaction kinetics relevant to real-world ambient exposure. To bridge this gap, we developed a high-temporal-resolution machine learning model based on XGBoost, trained on a comprehensive dataset comprising over 60,000 kinetic data points of ·O₂^- generation. The model demonstrated robust predictive performance (R² > 0.8) on the testing set, proving its capability to capture complex, non-linear kinetic patterns. Specifically, the model successfully reproduced the non-monotonic inverted V-shaped dose-response of Isoprene SOA and 9,10-Phenanthrenequinone (PQN) and accurately captured the antagonistic effects in PQN-Fe²⁺ mixtures, distinguishing these complex interactions from simple additive effects. Ongoing work focuses on applying this validated model to extrapolate cell-based kinetic data to environmentally relevant scenarios in human respiratory tract. We will first calculate the pollutant burden across different respiratory regions (e.g., trachea, bronchi, alveoli) by integrating ambient PM concentration data with a lung deposition model. We will then simulate region-specific ·O₂^- generation profiles by incorporating varying cell densities and analyzing kinetic parameters. Ultimately, this study aims to develop a model that translates ambient data into physiologically relevant oxidative profiles, providing a precise and cost-effective strategy for screening region-specific respiratory health risks.