From data mining to mechanistic prediction: a study on drug-induced lung injury integrating FAERS, machine learning, and network toxicology.

Drug-induced lung injury (DLI) is a serious and potentially fatal adverse drug reaction that remains difficult to detect early and continues to challenge both drug development and clinical safety. Because current diagnosis relies largely on exclusion and lacks specific early warning biomarkers, there is an urgent need for an integrated strategy that enables both risk prediction and mechanistic interpretation. To address this unmet need, we for the first time develop and validate a comprehensive integrated research strategy that combines FAERS database mining, XGBoost-based machine learning, and network toxicology, and is specifically designed for early DLI risk assessment and in-depth exploration of the underlying mechanisms. DLI-related adverse event reports were systematically extracted and standardized from FAERS to construct a curated small-molecule drug dataset. An optimized XGBoost model was then built using structural, physicochemical, and target-based features to predict DLI risk. In parallel, network toxicology was applied to construct drug-target-pathway networks for high-risk compounds and to identify key toxicological mechanisms. Our results highlighted MMP9 and ERBB2 as core targets associated with the pulmonary toxicity of Sunvozertinib and Zongertinib, and subsequent molecular docking and molecular dynamics simulations further suggested stable binding between the compounds and the above targets. This integrated framework enables efficient early prediction of DLI risk while providing mechanistic insights into drug-induced pulmonary toxicity. Overall, our study offers a practical and reproducible computational strategy for DLI risk assessment in drug development and supports safer clinical medication and rational drug design.