From phenotypic diversity to genotype selection: Integrating trait relationships and machine learning for high-yield Quinoa

Quinoa ( Chenopodium quinoa Willd.) possesses substantial phenotypic diversity, but the identification of high-yielding genotypes from large germplasm collections remains challenging because yield is determined by multiple interacting traits. This study evaluated the agro-morphological diversity of 142 quinoa genotypes and investigated the phenotypic profile associated with high-yield performance using multivariate and machine-learning approaches. Hierarchical cluster analysis classified the genotypes into five distinct groups, confirming substantial phenotypic diversity within the panel. Growth traits showed strong correlations across developmental stages; however, most vegetative traits were weakly associated with individual grain yield (IGY). Higher IGY was primarily associated with greater leaf area (LA) and thousand-seed weight (P1000), whereas excessive late plant height (PH) and number of branches (NoB) showed a negative relationship with IGY. Phenotyping across developmental stages improved discrimination between high-yield and other genotypes, and the complete predictor set including P1000 provided the strongest performance. Among the evaluated classifiers, including Logistic Regression, Linear Discriminant Analysis (LDA), Support Vector Machine with a radial basis function (RBF) kernel, and Random Forest, LDA achieved the highest PR-AUC and ROC-AUC, whereas Logistic Regression attained the highest balanced accuracy and identified a larger proportion of the observed high-yield genotypes. The high-yield phenotype was characterized by greater canopy development and chlorophyll status near 58 DAS, favorable seed mass, and moderate plant stature and branching. Consensus ranking further prioritized both observed high-yield genotypes and phenotypically similar candidates for subsequent evaluation. These findings demonstrate that integrating large-scale phenotypic characterization with multivariate and machine-learning analyses provides a useful framework for germplasm screening and candidate prioritization in quinoa breeding, although independent and multi-environment validation is required before predictive application.