Abstract
The global demand for improved productivity, sustainability, welfare, and quality in livestock production presents significant challenges for breeders. Understanding trait correlations, often driven by pleiotropy, is essential for simultaneously improving traits of economic interest. Integrating multi-omics data and functional annotations can improve the disentangling of biological processes underlying the pleiotropic effect. Network-based machine learning (ML) models offer a robust solution for this integration. This study estimated gene-level P-values for pleiotropic effects using two phenotypic datasets: (i) Trait_GWAS, with phenotypic values of 12 traits covering milk production, composition, cheeseability, and mastitis resistance; and (ii) EBV_GWAS, with estimated breeding values for five similar traits, excluding cheeseability. Weighted gene co-expression networks (WGCNs) were constructed from milk somatic cell transcriptomics of Assaf ewes. Gene-term networks were built from gene ontology, metabolic pathways, and quantitative trait loci annotation for the genes in the WGCN. These networks were processed through a representative learning pipeline to create a latent vector representing gene importance. A hierarchical model integrated gene-level P-values and the latent vector, generating posterior probabilities of association for each gene. Significant results included 14 and 111 genes for Trait_GWAS and EBV_GWAS, respectively, with three shared genes (PHGDH, SLC1A4, and CSN3). Prioritized genes were linked to biological processes such as amino acid transport, lipid metabolism, mammary gland development, and immune regulation, often involving multiple biological functions. This reinforces the potential pleiotropic role of these genes. These findings highlight the utility of network-based ML models for prioritizing candidate genes with pleiotropic effects on milk, cheese, and health-related traits in dairy sheep.