274
Prediction of Porcine Reproductive and Respiratory Syndrome Virus Serum Viral Level Phenotype from Gene Expression Profiles
Porcine reproductive and respiratory syndrome (PRRS) has been elusive to eradicate. The high morbidity and mortality associated with PRRSV infections has estimated associated annual costs over $664 million. Viral resistance has been documented in some pig breeds, however the genetic mechanisms involved are not fully understood. The objective of this study was to use gene expression profile data for prediction of animal-specific virus resistance status at early stages of development. Predictive ability of models that estimate viral load as a function of gene expression was evaluated using crossbred pigs infected with PRRS virus from the PRRS Host Genetics Consortium (PHGC). RNA and serum viral levels were obtained for 109 PHGC pigs at 4 and 7 days post infection (DPI). Transcriptional profiling was performed using the 70-mer 20K Pigoligoarray. The 4 and 7 DPI serum viral level phenotypes were regressed on genomic markers using three prediction methods: 1) a whole transcriptome linear method (genomic best linear unbiased prediction; GBLUP), 2) a linear regression method using forward stepwise variable selection with cross validation (FSVS) across all genes on the microarray (19,947 transcripts), and 3) a FSVS based only on significant genes from a previous differential expression analysis (518 transcripts for 4 DPI and 424 transcripts for 7 DPI). The GBLUP prediction accuracy was extremely low under 5-fold and 109-fold cross validation at 4 DPI (r = 0.09 to r = 0.14), and 7 DPI (r = 0.03 to r = 0.02). This contrasts with the predictive ability of genome-wide SNP markers used in genomic selection likely because correlation between gene expression profiles of genes is very different from correlation due to linkage disequilibrium between markers and QTL. The FSVS method resulted in high prediction accuracy. With 5-fold cross validation, the FSVS selected 20 variables and resulted in correlations of 0.87 and 0.92 at 4 and 7 DPI, respectively. Restricting the set of genes to the subset of genes differentially expressed in response to viral infection as a new set of predictors using the FSVS method reduced prediction accuracy (r = 0.71 at 4 DPI and r = 0.55 at 7 DPI). The unrestricted selection under cross validation of genes for a liner predictor shows the best predictive accuracy, and further evaluation of this method is warranted to improve predictive ability of PRRS phenotype using high-throughput gene-expression profiling.
Keywords: Pig, PRRS, Prediction