Integrating predicted gene expression into genomic selection: an evaluation in partridge chickens under realistic breeding scenarios

Abstract Background Genomic prediction (GP) has enhanced breeding value prediction using genome-wide markers and genotype–phenotype associations. Integrating transcriptomic data, which are closer to phenotypes, may improve breeding value prediction accuracy. While multi-omics integration in livestock is cost-prohibitive, resources such as FarmGTEx enable large-scale SNP-based prediction of estimated breeding values (EBVs) for gene expression. However, utility of these predicted EBVs for enhancing GP accuracy remains underexplored. We evaluated the integration of observed gene expression and gene expression EBVs into genomic prediction models for live, dressed, and eviscerated weights in 350 Tiannong Partridge Chickens. Whole-genome resequencing and pectoral muscle transcriptomes were generated, and multi-omics models were assessed across three scenarios: (I) both reference and validation populations have transcriptomic data; (II) only the reference population has transcriptomic data; (III) neither has transcriptomic data—requiring cross-population prediction using the ChickenGTEx reference panel. We compared genomic–transcriptomic best linear unbiased prediction (GTBLUP) models incorporating gene expression EBVs predicted via best linear unbiased prediction (BLUP) or elastic net (ENET) (GTBLUP_pred_BLUP and GTBLUP_pred_ENET, respectively), genomic-feature BLUP (GFBLUP) based on expression quantitative trait loci, and standard genomic BLUP (GBLUP). Model prediction accuracies were compared using t -tests followed by Holm correction. Results In Scenario I, integrating observed expression data significantly improved phenotype prediction accuracy over GBLUP (adjusted p < 0.05; reference-only). Subsequent comparisons are genomic prediction accuracy. For Scenario II, the GTBLUP_pred_ENET model (using genes with h 2 > 0.7) showed modest numerical gains over GBLUP when using gene expression EBVs of the validation population, although the differences were not statistically significant (adjusted p > 0.05). Furthermore, using the cross-population ChickenGTEx reference panel for expression EBV prediction (Scenario III), GTBLUP_pred_ENET ( h 2 > 0.3) again yielded consistently higher prediction accuracies than GBLUP, albeit without statistical significance (adjusted p > 0.05). In contrast, GFBLUP showed minimal or negative gains, especially when regulatory signals were extrapolated from external populations. For expression EBV prediction, ENET generally outperformed BLUP in Scenarios II and III. We found a low-to-moderate positive correlation between gene expression prediction accuracy and its cis-heritability, although overly stringent heritability filtering reduced genomic prediction accuracy. Conclusions These results highlight the potential of predicted expression EBVs for GP, pending validation in larger, more diverse populations and traits, alongside methodological refinements.