MGIDI and Explainable Machine Learning for Multi-Trait Selection and Yield Drivers in Rainfed Bread Wheat

Identifying superior wheat genotypes under rainfed semi-arid conditions and understanding the traits that drive yield remain challenging, and multi-trait selection indices and explainable machine learning are seldom used together. This study combined the Multi-Trait Genotype–Ideotype Distance Index (MGIDI) with explainable machine learning to evaluate 20 bread wheat (Triticum aestivum L.) genotypes for 14 phenological, physiological and yield-component traits in a randomised complete block design in a single rainfed season in Diyarbakır, Türkiye. Broad-sense heritability was high for all traits (H2 = 0.84–0.99), and the close agreement between genotypic and phenotypic coefficients of variation indicated a predominant genetic contribution to phenotypic variation. Grain yield, the number of grains per spike, and grain weight per spike showed the highest expected genetic advance (GAM = 17–22%). MGIDI-based selection of the top 25% of genotypes increased grain yield by 9.50%, with concurrent gains of 6–13% in spike length, spikelets per spike and grain weight per spike, and the relationship between MGIDI and grain yield was negative and significant (R2 = 0.58, p < 0.001). Among five learners evaluated under block-stratified GroupKFold cross-validation, Gradient Boosting performed best (CV R2 = 0.740; training R2 = 0.998; RMSE = 30.55 kg da−1; MAE = 24.43 kg da−1). SHAP analysis identified grain weight per spike, plant height and days to flowering as the main yield drivers, and these rankings were corroborated by model-independent permutation importance. Combining MGIDI with explainable machine learning allowed genotype selection and yield-determining traits to be evaluated within a single framework, and the selection decisions were biologically consistent.