Integrative metabolomics and machine learning reveal diagnostic biomarkers for gelsenicine intoxication

Gelsemium elegans Benth. (G. elegans) is a highly toxic medicinal plant traditionally used to treat pain and inflammatory disorders. Accidental ingestion or misuse often causes severe poisoning and death, primarily due to respiratory depression. Its most toxic alkaloid, gelsenicine, undergoes rapid metabolism in vivo. The lack of specific biomarkers hinders timely and accurate assessment of poisoning severity.

This study combined untargeted metabolomics with machine learning (ML) to distinguish gelsenicine-induced fatal intoxication (GFI) from multiple hypoxia-related deaths in mice. Key discriminatory metabolites were identified, and a serum-based classification model was established and validated to support precise clinical diagnosis and forensic identification.

Serum samples were collected from GFI and three non-drug-related deaths (NDRD) groups: cervical dislocation (CD), compressive asphyxia (CA), and asphyxia due to ambient-hypoxia (ADAT). Untargeted metabolomic profiling was performed using UPLC-HRMS. Differential metabolites were identified with Compound Discoverer and SIMCA software. Biomarker selection and model construction were conducted on the MetaboAnalyst platform using random forest (RF)-based receiver operating characteristic (ROC) analysis. Model performance was evaluated across multiple ML algorithms. Sensitivity was tested with GFI samples at 1, 2, and 4 mg/kg doses, and specificity was assessed against three other neurotoxicant-related fatal intoxications (ONDFIs): isoflurane (IFI), carbon monoxide (COFI), and methamphetamine (MFI).

Metabolomic analysis revealed significant differences between GFI and NDRD groups. Three key metabolites, creatinine, valylserine (Val-Ser), and tyrosyl-phenylalanine (Tyr-Phe), were selected to develop a classification model. The model showed promising predictive performance with area under the curve (AUC) values above 0.9 across multiple algorithms. It accurately identified GFI across different exposure doses and distinguished GFI from ONDFIs with reasonable accuracy. Further targeted metabolomics analysis suggested that creatinine levels exhibited dose-dependent changes. Overall, the model demonstrated satisfactory sensitivity and specificity.

By combining untargeted metabolomics with ML, a classification model based on creatinine, Val-Ser, and Tyr-Phe was established. The model effectively distinguishes GFI from multiple hypoxia-related deaths and demonstrates high accuracy, sensitivity, and specificity, indicating its potential for forensic precision identification and clinical diagnosis of gelsenicine poisoning.