Discovery of Q-Markers in traditional Chinese medicine based on machine learning: Application and verification in Xiaoyao Heji

Objective: Identification of quality marker (Q-Marker) is central to the modernization and quality control of traditional Chinese medicine (TCM). Existing methods, though diverse, lack a unified and scalable strategy to address the chemical and pharmacological complexity of TCM prescriptions. This study established a machine learning-driven framework to integrate multi-source data and systematically prioritize Q-Marker, using Xiaoyao Heji (XYHJ), a modern antidepressant formulation derived from Xiaoyao San, as a case study. Methods: Firstly, a dataset of 302 compounds was constructed by summarizing 14 reported Q-Marker discovery strategies and annotated with five core dimensions-compatibility, specificity, measurability, traceability, and effectiveness. Subsequently, five ranking models (RankNet, Random Forest, XGBoost, Linear Regression and K-Nearest Neighbors) were trained with 8-fold GroupKFold cross-validation and evaluated by normalized discounted cumulative gain (NDCG), mean average precision (MAP), and Recall. Thirdly, the optimal model was applied to XYHJ, followed by systematic characterization of chemical profiles, quantitative analysis, drug metabolism and pharmacokinetics (DMPK), quality transitivity, and bioactivity. Finally, Q-Markers were further validated through radar chart modeling, Monte Carlo simulations, and chemometric analyses. Results: Through a comparison of the machine learning models, XGBoost outperformed other models, exhibiting high and stable performance with NDCG, MAP, and Recall values ranging from 0.84 to 0.95. For the first time, the XYHJ prescription was comprehensively evaluated across its chemical profile, DMPK properties, quantitative content, manufacturing traceability, and biological efficacy. Using the optimal XGBoost ranking model, 15 components were screened and ranked, successfully identifying paeoniflorin, albiflorin, and glycyrrhizic acid as critical Q-Marker. The reliability of the ranking approach was further validated by integrating a previously published radar chart model with Monte Carlo simulations with over 10 000 sampling iterations. Furthermore, fingerprint similarity and chemometric analysis demonstrated the effectiveness of these Q-Markers in distinguishing XYHJ produced through different manufacturing processes. Conclusion: This study developed a machine learning-driven ranking framework that unifies multiple Q-Marker discovery strategies. Using XYHJ as a case study, this approach not only enhances its quality control system but also provides a valuable reference for the standardization and modernization of other TCM prescriptions.