Incorporating genomic and transcriptomic effects in joint linear and non-linear structural models for predicting complex traits in pigs

Phenotypes in livestock are shaped by genetic variation as well as downstream regulatory mechanisms, making the prediction of complex traits a key challenge for animal breeding. Transcriptomic data represent an intermediate biological layer between genotypes and phenotypes and may capture regulatory signals not fully explained by genomic information alone. The objective of this study was to evaluate the contribution of blood transcriptomic data, alone or combined with genomic information, to predict six immune, stress, and production traits in 255 Duroc pigs. Four traits were closely related to the sampled tissue and timepoint, whereas two were less biologically relevant. Bayesian regression methods (BayesC and RKHS) and a neural network linear mixed model were compared using either all transcripts or subsets selected by Partial Least Squares (PLS). High prediction accuracy was obtained for immunity-related traits, such as gamma delta T cells and leukocyte counts, with correlations of 0.74 and 0.67, respectively, when transcriptomic data were used. Moderate improvements were observed for cortisol prediction (r = 0.39), whereas SNP-based models performed best for carcass weight (r = 0.45). PLS-based feature selection showed that a small subset of features can perform equally well or better than the whole transcriptomic dataset and identified biologically relevant candidate genes, including MAF, SOX13, DDIT4, and FOS. In conclusion, blood transcriptomic data substantially improved prediction performance for traits biologically related to the sampled tissue, whereas SNP-based models performed better for less relevant traits, and combining omics provided only modest and non-significant gains; feature selection was essential to enhance prediction performance, computational efficiency, and to facilitate the identification of immune-related candidate genes.