Machine learning-based prediction of dynamic height heterosis with pathway biomarkers in rice

The development of robust biomarkers enables accurate prediction of complex phenotypes. However, the dynamic nature of biomarkers is often underestimated since their quantitative changes during development are directly connected to phenotypic transformations, influencing both crop agronomic traits and human diseases. Here, we performed network analysis of untargeted metabolite profiles to investigate height heterosis in rice, which is dynamic that varies during development and is a key determinant of yield heterosis. We found that the levels of pyruvaldehyde were predictive of height heterosis specific at the seedling stage, while 4-hydroxycinnamic acid positively correlated with height heterosis across four developmental stages. We identified metabolic pathways associated with height heterosis and found that metabolomic changes during the elongation stage had a greater impact than those in other stages. Finally, 11 heterosis-associated pathways were developed into metabolomic biomarkers through random forest analysis, successfully predicting height heterosis in an independent population under different growth conditions. This study elucidates the metabolomic landscape of dynamic height heterosis in rice and develops pathway biomarkers for complex phenotypes, demonstrating robustness across diverse populations, environments, and developmental stages.