Genomic prediction in Persian walnut: Optimization levers according to genetic architecture of complex traits

Persian walnut (Juglans regia L.) is a widespread cultivated nut tree species in temperate regions. Advances in genomic tools, such as the high-density Axiom J. regia 700K single nucleotide polymorphism (SNP) genotyping array, enable the exploration of genomic prediction (GP) for this crop. This study is the first to evaluate GP accuracy and several influencing factors in walnut for traits related to phenology and nut quality. A core-collection of 170 accessions was phenotyped for 25 traits over 1 or 2 years. Highly heritable traits, such as budbreak date and female flowering date, were predicted with high accuracy (∼0.75) using ridge regression best linear unbiased prediction (rrBLUP). Three key factors influencing GP performance were examined: marker density, prediction model, and training set size. Selecting the top 1% of 364,275 SNPs based on their variance (∼3600 SNPs) was sufficient to achieve accurate predictions. Bayesian models slightly improved prediction accuracy for some traits when using this reduced SNP set, but rrBLUP provided robust results, balancing accuracy, simplicity, and computational efficiency. Training population size also influenced accuracy, with a subset comprising 50% of the population still yielding reliable predictions. Optimization of training set was assessed using coefficient of determination mean, prediction error variance mean, and mean relatedness (MeanRel) parameters, with MeanRel performing best for shell traits. However, incorporating SNPs identified in genome-wide association study into the prediction models did not enhance accuracy. In summary, this study demonstrates the feasibility and potential of GPs for walnut breeding programs using a core-collection, offering valuable insights for optimizing GP approaches in this crop.