Existing Bayesian genomic evaluation methods such as BayesRC can account for biological annotations but not more than one annotation per SNP. However, when a SNP is associated with multiple sources of functional annotation, it is not straightforward to identify the most relevant annotation for a trait of interest. To handle multi-annotated SNPs, the BayesRCπ approach assigns a multi-annotated SNP to its optimal annotation and within a specific effect class (null, small, medium or strong effect). These strategies were tested in two dairy breeds (Montbéliarde and Normande) and one beef breed (Charolaise). Four different annotation classes were considered (50K chip, GWAS, GWAS meta-analysis, genomic features) for a total of around 100,000 and 50,000 SNPs in the dairy and beef breeds, respectively. The traits of interest were milk, protein, and fat yields, protein and fat contents in dairy cattle, and weight at 18 months, thickness of bones, muscular and skeletal development in beef cattle. Although assuming different variance priors in a BayesR model led to a significant improvement in the accuracy of genomic predictions for a number of traits, incorporating annotation classes via the BayesRC and BayesRCπ models did not result in any additional gain. However, the posterior distribution of SNPs in the different effect classes (null, small, medium or strong) strongly differed between models, with the SNPs having a strong biological annotation being more frequently assigned to the medium and strong effect classes with the BayesRCπ model than with the BayesR model. This suggests that the biological information is useful in identifying SNP with strong effects, which may favour more robust prediction equations over time.