Identification of potential biomarkers for Polygonum multiflorum-induced liver injury in a clinical cohort: integrating machine learning of metabolomics with transcriptomic profiling.

ETHNOPHARMACOLOGICAL RELEVANCE: Polygonum multiflorum Thunb. (PM), synonymous with Reynoutria multiflora (Thunb.) Moldenke is known as He Shou Wu in traditional Chinese medicine. Despite its long history of use, PM has been associated with drug-induced liver injury (PM-DILI), a rare but potentially serious adverse event. Therefore, identifying reliable biomarkers is crucial for the early detection of PM-DILI, risk prediction, and the safe and rational clinical use of PM. AIM OF THE STUDY: To identify candidate diagnostic biomarkers for PM-DILI and obtain preliminary mechanistic insights by integrating metabolomic and transcriptomic profiling of a well-matched clinical cohort. METHODS: A case-control study of 50 patients with PM-DILI and an equal number of PM-tolerant (PM-T) controls was conducted. Plasma metabolites were profiled using untargeted liquid chromatography-mass spectroscopy analysis. The cohort was stratified split into a training set (70%) and an independent test set (30%). A high-confidence candidate pool was generated within the training set by intersecting metabolites from Weighted Gene Co-expression Network Analysis (WGCNA)-derived clinical trait-associated modules with differential metabolites. Feature selection was performed using a random forest-based recursive feature elimination (RFE) approach. The selected features were then evaluated using three machine learning algorithms, Least Absolute Shrinkage and Selection Operator (LASSO), Support Vector Machine (SVM), and Random Forest (RF), to identify candidate biomarkers. Transcriptomic and coexpression analyses were performed in a subset (12 per group), with key findings were further supported by reverse transcription-quantitative polymerase chain reaction (RT-qPCR). RESULTS: The PM-DILI and PM-T groups were well-matched at baseline. Integration of WGCNA with differential analysis yielded a high-confidence metabolite pool. Following feature selection and evaluation by the three machine learning algorithms, cholesterol sulfate was identified as a candidate biomarker. Notably, in the LASSO and SVM models, cholesterol sulfate demonstrated outstanding discriminatory power, with area under the curve values of 0.80, sensitivity of 0.5333-0.6667, and specificity of 0.8667-1.0 in the independent test set. Transcriptomics identified coexpressed genes (LINC02528-RAB39A), validated using RT-qPCR, whose expression correlated with the metabolite, suggesting a candidate multi-omics association in PM-DILI. Given the relatively small transcriptomic sample size, these results should be interpreted as exploratory and require further validation. CONCLUSION: Based on multi-omics analysis of clinical plasma samples, cholesterol sulfate was identified as a candidate diagnostic biomarker, and a correlated gene pair LINC02528-RAB39A was highlighted as a candidate transcriptional signature associated with PM-DILI.