An Interpretable Ensemble-Learning Framework for Decoding Multi-Component–Multi-Bioactivity Relationships in Salvia miltiorrhiza Based on 1H NMR Metabolomics and Lifecycle-Oriented Bioassays

Background: Linking metabolite composition to biological activities remains challenging in herbal medicines, where diverse metabolites may collectively contribute to pharmacological effects. Although metabolite profiling is fundamental to quality assessment, conventional quality evaluation often provides limited information on how variation in metabolite composition relates to biological activities. Objectives: This study aimed to develop a multi-component–multi-bioactivity framework integrating metabolite profiling and bioactivity assessment, enabling bidirectional mapping between metabolite composition and biological responses and providing a functional perspective for quality characterization of complex herbal medicines. Salvia miltiorrhiza Bunge (Danshen), a widely used medicinal plant with a chemically diverse metabolome and well-documented cardiovascular activities, was selected as a representative model system. Methods: Quantitative 1H NMR spectroscopy was employed to profile 28 metabolites in Danshen, while zebrafish-based bioassays were used to assess three cardiovascular-related bioactivities: pro-angiogenic, anti-thrombotic, and vasodilatory activities. A lifecycle-oriented approach was introduced for bioassay development to improve the robustness and reproducibility of biological measurements. Multiple machine learning approaches were integrated within a two-level stacking ensemble framework to construct quantitative metabolite–bioactivity relationship models and capture nonlinear associations between metabolite composition and biological activities. Results: The stacked models consistently outperformed the individual learners across the three cardiovascular bioactivities, achieving R2 values of 0.895–0.931 and RMSE values of 0.116–0.130. External validation using 10 independent Danshen samples further supported the predictive reliability of the stacked models. Based on model-derived metabolite contributions, a quantitative and interpretable metabolite–bioactivity network linking 28 metabolites with three cardiovascular-related activities was constructed, revealing both shared and activity-specific metabolic contributors. Conclusions: This study establishes an interpretable framework integrating NMR-based metabolite profiling, robust multi-endpoint bioassays, and ensemble learning to connect metabolite composition with functional bioactivity. The framework provides a complementary bioactivity-informed perspective for the quality characterization of medicinal plants and supports systems-level evaluation of complex herbal medicines.