Integrated serum metabolomics and machine learning for classification of Datura stramonium L. exposure using plant part-specific metabolic signatures.

Datura stramonium L. (DS) poisoning remains a significant toxicological concern in clinical and forensic contexts. In forensic casework, leaf ingestion is mainly associated with accidental exposure, whereas seed ingestion is more frequently linked to suspected intentional poisoning. However, conventional toxicological analysis based on tropane alkaloids cannot distinguish between these exposure sources, limiting forensic interpretation. This study developed a serum metabolomics strategy integrating machine learning to viably discriminate between seed- and leaf-derived Datura exposure using metabolic signatures. Serum samples from DSS (seed-derived) and DSL (leaf-derived) exposure models were analyzed using UPLC-Orbitrap-MS-based untargeted metabolomics. Multivariate analysis revealed distinct metabolic profiles between DSS and DSL groups, involving alterations in alkaloids, phenylpropanoids, lipids, and organic acids. Later, a supervised machine learning framework was developed to classify exposure sources based on metabolic signatures. The optimized model achieved an area under the curve (AUC) of 0.90. Feature interpretation further identified key metabolites contributing to classification, enabling biological interpretation of forensic decision boundaries. Meanwhile, higher metabolites in machine learning can also reflect different pathological mechanisms, providing a mechanistic basis for exposure differentiation. These results demonstrate that the integration of serum metabolomics and machine learning enables accurate and interpretable forensic source attribution of Datura poisoning. The proposed framework extends traditional toxicological analysis from exposure confirmation to data-driven forensic reconstruction, offering a novel approach to high-resolution forensic toxicology in complex plant poisoning cases.