Our objective was to investigate the potential benefits of using sequence data to improve across breed genomic prediction, using data from five French and Danish dairy cattle breeds. First, QTL for protein yield were detected using high density genotypes. Part of the QTL detected within breed was shared across breed. Second, sequence data was used to quantify the loss in prediction reliabilities that results from using genomic markers rather than the causal variants. 50, 100 or 250 causative mutations were simulated and different sets of prediction markers were used to predict genomic relationships at causative mutations. Prediction of genomic relationships at causative mutations was most accurate when predicted by a selective number of markers within 1 Kb of the causative mutations. Whole-genome sequence data can help to get closer to the causative mutations and therefore improve genomic prediction across breed