Prioritizing steatogenic chemicals through integration ToxCast™ data, machine learning, and experimental validation

As hepatic steatosis driven by environmental exposures increasingly contributes to the global burden of metabolic disease, identifying and prioritizing high-potency steatogenic chemicals is critical for enabling risk-oriented toxicological and environmental regulation. Leveraging the well-established adverse outcome pathway framework for hepatic steatosis, we integrated ToxPi scores derived from 14 molecular initiating events in the ToxCast™ database with in vivo validation in zebrafish. This integrated approach enabled the construction of a training set comprising chemicals with distinct steatogenic potency. Feature selection via Kruskal-Wallis test identified 11 key bioassays, with OT_FXR_FXRSRC1_0480 and NVS_NR_hGR contributing most to model performance. Using leave-one-out cross-validation, the SVM model achieved 91.7% accuracy in the training set. External validation on 35 compounds, although based on binary activity labels, resulted in 77.1% accuracy, indicating moderate but promising generalizability. Final predictions on 345 curated ToxCast™ chemicals (from a total of 9924) were categorized as high- (37.97%), moderate- (18.84%), and null-effect (43.19%) on steatogenic potence by Random Walk with Restart algorithm. In vivo validation of 14 predicted compounds confirmed the model's robustness, and in vitro lipid staining assays in HepG2 cells further demonstrated concordance. This study revealed that several emerging contaminants, including isodecyl diphenyl phosphate, 3,3'-dimethylbisphenol A, tetrabutyltin, tetrabromobisphenol A bis(2-hydroxyethyl) ether, trixylyl phosphate and quinoxyfen, exert high steatogenic potency. These findings underscore the utility of integrating high-throughput data with predictive modeling and experimental validation to prioritize high-potent steatogenic chemicals.