Metabolomic signatures predict heterotic performance in opium poppy (Papaver somniferum)

Metabolomics-based prediction models are reshaping hybrid breeding by enabling early and accurate identification of superior genotypes. Papaver somniferum (opium poppy), a pharmaceutically important crop producing benzylisoquinoline alkaloids, is particularly suited to benefit from this approach. Although metabolomics and machine learning have been widely applied in food and oilseed crops, their potential remains largely unexplored in medicinal plants, especially opium poppy. Here, we developed a metabolomics-guided, machine learning-assisted framework to predict hybrid performance and accelerate breeding progress. A cross between two genetically divergent parental lines, Sujata (low alkaloid) and Thial (high alkaloid), yielded 233 F₁ hybrids evaluated at physiological maturity for five key alkaloids and three yield-related traits. Principal component analysis separated phenotypic variation into chemical and agronomic axes, revealing distinct hybrid clusters. Six machine learning algorithms (Multiple Linear Regression, Ridge Regression, LASSO, Random Forest, XGBoost, and Support Vector Regression) were applied under three predictive scenarios: metabolomics-only (MP), morphology-only (MMP), and combined metabolomics-morphology (MM_MP). Random Forest, XGBoost, and LASSO consistently achieved the highest prediction accuracy, particularly under the MP and MM_MP frameworks. Noscapine (R² = 0.654) and morphine (R² = 0.611) emerged as the most reliably predicted traits. Clustering analyses identified elite alkaloid-rich chemotypes (e.g., H83, H181) and dual-purpose ideotypes (e.g., H221, H78) that combined high alkaloid content with strong biomass yield. Feature importance analysis highlighted biosynthetic relationships among alkaloids, offering mechanistic insights into trait regulation. Overall, this study demonstrates a scalable metabolomics-assisted prediction framework that enables first-cycle ranking of F₁ hybrids from mature-stage data, thereby supporting earlier decisions without assuming cross-stage predictability.