Machine learning–assisted early selectition of S1 segregants in strawberry, Fragaria × ananassa Duch

Abstract

Strawberry (Fragaria × ananassa) breeding requires analytical tools that integrate productivity and fruit quality in highly variable segregating populations. This study evaluated the integration of mixed models and machine learning algorithms to support early selection of S1 genotypes with high yield potential and fruit quality. A total of 320 genotypes were evaluated under an augmented block design for phenological, vegetative, productive, and fruit quality traits. Broad-sense heritabilities and predicted genetic values were estimated using mixed models and integrated with predictions from XGBoost and CatBoost. Moderate to high heritabilities were observed for structural and productive traits (H² = 0.52–0.86). Positive genetic correlations were found between fruit number and yield, whereas soluble solids content was negatively associated with yield components, indicating trade-offs between productivity and fruit quality. XGBoost showed high predictive accuracy for commercial yield (R² = 0.908), intermediate performance for fruit weight (R² = 0.461), and poor performance for soluble solids content as a continuous variable (R² = −0.096). Recoding soluble solids content into categorical classes improved interpretability, although class discrimination remained limited. Selection indices based on best linear unbiased predictors (BLUPs), machine learning predictions, and their combination produced consistent rankings and identified genotypes with balanced agronomic and quality performance. Integrating mixed models and machine learning into multivariate selection indices provides a useful strategy for improving early selection efficiency in strawberry breeding programs.