Abstract
Introduction Pearl millet, dubbed a “Nutri-cereal”, has a high content of protein, starch, fiber, mineral and fatty acids. Its resilience in adverse agro-climatic conditions sets it apart from major cereals. Despite this, understanding how its genetic diversity affects physiological traits and metabolic responses remains limited. Predictive metabolomics, merging metabolomics with artificial intelligence, allows for the comprehensive top-down modelling —from phenotype to the mechanism— of various phenotypic traits.Objectives To discover predictive biomarkers for phenotypic traits in the Brazilian germplasm core collection of 203 genotypes of pearl millet through the combination of predictive metabolomics with machine learning.Methods Untargeted metabolomics was conducted using UHPLC-LTQ-Orbitrap-HRMS to obtain metabolite profiles, from the central and specialised metabolism of the pearl millet core collection. Generalised linear modelling with penalisation (GLMNET) was applied to explore the correlation between metabolism and phenotypic traits.Results Our model successfully predicted eight qualitative traits from the pearl millet core collection, with accuracy ranging between 74% and 87%. From, 834 potential unique biomarkers (575 annotated-ion features and 259 unknowns) have been annotated as top metabolic predictors. It is noteworthy that the majority of the top metabolic predictors were from the carbohydrate, amino acid, flavonoid, and terpene subclasses.Conclusions This is the first report on leveraging a germplasm bank of pearl millet for metabolome characterisation and subsequent predictive modelling of important agronomic traits. These outcomes hint at the robustness of employing GLMNET for predicting metabolic biomarkers crucial in selecting genotypes for future breeding programmes.