Comparative Genomic Prediction in Winter Wheat

Abstract

Haplotype blocks offer potential advantages over single SNP markers in genomic prediction by capturing local epistatic interactions and ancestral relationships. However, their effectiveness varies by trait and block construction method. This study evaluates genomic prediction for 11 traits in winter wheat using haplotype blocks constructed based on linkage disequilibrium (LD), fixed SNP numbers, fixed lengths in cM, and HaploBlocker, alongside single SNP markers. Data from 361 genotypes and single-year field trials were analyzed using ridge regression best linear unbiased prediction (RR-BLUP), restricted maximum likelihood ANOVA (RMLA), and genomic and variance component based on haplotype blocks (GVCHAP) in a cross-validation study. LD-based blocks showed the highest prediction accuracies for resistance traits (Blumeria graminis, Puccinia triticina, and Fusarium graminearum), while HaploBlocker outperformed other methods for protein concentration and resistances to Septoria tritici and P. striiformis. Machine learning methods were also assessed for their potential to improve prediction accuracy and resistance genotype identification. Nineteen prediction approaches, combining different models (RR-BLUP, RMLA, Bayesian ordinal regression, support vector machine, gradient boosting machine, and random forest), predictors (SNPs, LD-based blocks, autoencoder variables, and GWAS-selected SNPs), and response values (untransformed and logit transformed scores), were evaluated using cross-validation and Cohen’s κ. While RR-BLUP consistently performed well, gradient boosting machine and random forest achieved higher accuracy for P. triticina (0.71 vs. 0.64 for RR-BLUP). Cohen’s κ was used to assess the ability of each method to correctly identify the most resistant genotypes, demonstrating that machine learning models offered improvements in resistance identification, especially for P. triticina. A major finding is that no single method or predictor set universally outperforms others. The choice of model and predictors must be tailored to the genetic architecture of the trait being studied. Logit transformation proves particularly useful for resistance scores, ensuring that predicted genotypic values remain within an interpretable range. This thesis underscores the importance of a trait-specific approach to genomic prediction, emphasizing that effective prediction relies on selecting models and predictors that align with the unique characteristics of the traits.