Genomic Prediction in Quinoa Across Contrasting Environments Using Statistical and Machine Learning Models

Quinoa (Chenopodium quinoa Willd.) is gaining global importance for its nutritional value and adaptability; however, breeding progress remains limited. Genomic selection (GS), combined with rapid generation cycles, offers a strategy to accelerate genetic improvement. We conducted whole-genome resequencing of 610 accessions and present the first evaluation of genomic prediction in quinoa evaluated across six field trials in Australia and Pakistan for seven phenological and yield-related traits. Using ∼1.8 million single-nucleotide polymorphisms, we compared four models-genomic best linear unbiased prediction, reproducing kernel Hilbert space, BayesC, and light gradient boosting machine-for genotype ranking under four cross-validation schemes: predicting new genotypes (CV1), sparse testing (CV2), leave-one-location-year-out (CV0), and across locations. Model performance was evaluated using Pearson's correlation for overall accuracy and normalized discounted cumulative gain (NDCG@10) for ranking top performers. The four models were similar, with no method dominating across traits. NDCG@10 scores revealed that predictions remained useful for selecting superior genotypes even for difficult traits. Prediction accuracy was strongly associated with heritability and trait correlations across and within- location environments. Accuracy was highest for developmental traits and lowest for seed yield, while seed traits showed location-specific responses with higher accuracy in Australia. These findings support GS as a promising tool for quinoa breeding and provide benchmarks for global implementation.