DNA of each accessions was digested with ApekI. Bar-coded DNA of 96 accessions was pooled and sequenced with a Genome Analyzer II. The short reads were aligned to BTx623 reference genome v1.4 with BWA. SNPs were called with the GBS pipeline in TASSEL. After filtering out SNPs with missing data rate >80% and minor allele frequency < 0.5%, missing SNP calls were imputed with fastPHASE. Then the whole 962 panel was split into two dataset, one with 229 accessions as the training set and another with 663 as the validation set. The training set was extensively phenoytped in multiple years and locations and used to build genomic prediction model. The genomic prediction models were then used to predict the trait values for the 663 validation set. Then a subset of 200 accessions from the 663 validation set was planted to test the prediction model. The observed phenotypes are the BLUP estimates from multiple replications, years and locations.