The tremendous increase in throughput of genotyping techniques opened appealing perspectives for genome-wide selection (GWS), which could enhance breeding efficiency by decreasing generation interval and increasing selection intensity and/or accuracy of breeding values. In GWS a large training population with both phenotypic and genotypic data is used to construct a statistical prediction model which is then applied to estimate Genomic Breeding Values (GBV) of individuals that only have genotypic data. In the EU-FP7 project FruitBreedomics, we performed a pilot study of GWS in apple. The two main objectives were to provide proof of principle and to evaluate the accuracy of prediction with respect to the relatedness between the test and the training populations. Hereto the phenotypic means of the 50 best and 50 worst predicted individuals of a test progeny comprising 700 individuals were compared and correlations between predicted GBV and phenotypic data were compared for smaller full-sib families of different relatedness.The training population included 20 full-sib families comprising 992 individuals genotyped with an Illumina 20K SNP array and phenotyped for fruit quality traits. The test population comprised four progenies from commercial breeding programs totaling 1500 individuals that were genotyped with 512 SNP that had been selected among the 20K for heterozygosity in the parents of the test population and genome-wide coverage. SNP genotypes were completed through imputation and fed into the calibrated prediction model to obtain GBV on the 1500 offspring. We will present the results and discuss challenges remaining to bring GWS into practice in Rosaceae species.