Enhancement of QTL Mapping and Genomic Prediction for Bovine Respiratory Disease Using Sequence Imputation and Feature Selection

Genomic selection and GWAS have proven successful for livestock research and breeding. Expanding the application of genomics to novel traits with high economic importance such as disease resistance holds great promise. It is important to maximize the utility of data from research studies designed to create predictions for these traits, as routine retraining may prove difficult. Here we take two approaches to improve Bovine Respiratory Disease (BRD) genomic predictions: imputation to sequence and feature selection. From the original study of BRD, BovineHD genotypes for 2703 case-control Holstein animals observed in two states were imputed to the 1000 Bull Genome Project reference set, resulting in 11 million imputed variants with minor allele frequency >5%. Using these imputed data, our estimate genomic heritability rose from 13.6% to 14.4%. There was also a substantial increase in the resolution of the genomic regions detected as harboring QTL. In genomic predictions using BSLMM models implemented in GEMMA, the imputed data did not result in a substantial improvement in the accuracy of predictions. To address this issue, we applied a feature selection approach in which we selected two sets of variants: those in regions with differentially expressed genes from a BRD challenge experiment using RNA-Seq and variants in QTL regions identified in the imputed variant GWAS. Including only these variants in the model improved phenotype prediction by 50% from an 11% correlation between breeding values and phenotypes to a 17% correlation. Training using data from one state represented in the study and predicting in the other resulted in poor accuracies, but improved substantially with the use of feature selection. These methods hold great promise for applying genomic selection to livestock populations, particularly for difficult to measure traits.