Genomic Prediction for Powdery Mildew Resistance in Wheat: A Thirty-Model Benchmark of Conventional, Machine-Learning, and Deep-Learning Approaches

Powdery mildew, caused by Blumeria graminis f. sp. tritici (Bgt), is a major foliar disease of bread wheat (Triticum aestivum L.). Genomic prediction (GP) can accelerate breeding for this trait, but conventional, machine-learning (ML), and deep-learning (DL) methods have rarely been compared side by side. Using the Wheat Association Mapping Initiative (WAMI) panel (n = 286), phenotyped for powdery mildew severity over three years (DS1–DS3) and genotyped with the Illumina 90K SNP array, we compared 30 GP models under per-year and genotype-grouped pooled cross-validation. Broad-sense heritability ranged from 0.65 to 0.89. In the pooled, leakage-safe analysis, six Bayesian-alphabet models clustered at the top (r = 0.291–0.299), too close to identify a single best model; partial least squares (PLS) regression was the strongest non-Bayesian model, and elastic net was the strongest ML model within each year. Within-year accuracy under random 5-fold cross-validation was higher (up to r ~ 0.70 in DS2) but may be optimistic, as folds were not stratified by genomic relatedness. Accuracy showed limited, non-monotonic sensitivity to marker density and increased, though not strictly monotonically, with training population size. Leave-one-year-out (LOYO) accuracy was near zero (mean r = 0.009); fitting year as a fixed effect did not improve it, and cross-year genetic correlations were low (0.08–0.15), indicating limited shared genetic control across years. DL results apply only to the architectures and sample size tested. With pooled accuracy near r = 0.30, these models suit within-season prediction in this panel, whereas reliable cross-season application is not supported by the present results and will require broader multi-environment training and/or season-specific recalibration.