An Integrative Multi-Omics Framework for Genomic Prediction in Livestock Using Bayesian Deep Neural Networks

Lifeng XingHuihui WangYonglin Yang^*Bujun Mei

• Hetao College, Bayannur, China

Background: Genomic selection has accelerated genetic gain in livestock, yet most models rely solely on SNP genotypes and ignore regulatory layers that mediate genotype-to- phenotype translation. Objectives: We introduce an integrative framework that unifies whole-genome sequence variants, tissue-specific expression profiles, and curated QTL/GWAS annotations within a relational database, and couples these data with a Bayesian deep neural network (DNN) for genomic prediction in cattle. Methods: A 100 GB PostgreSQL schema links 20 M imputed SNPs, 15 k gene-expression traits, and >10 k bovine QTL. Feature-level biological priors are extracted on-demand and passed to a two-layer Bayesian DNN implemented in TensorFlow Probability. Performance is benchmarked against GBLUP, BayesA/RC, LASSO, Random Forest, and a non-Bayesian DNN using simulated and publicly derived data sets. Results: The integrative Bayesian DNN raised predictive correlation by 10–15 % over genotype-only baselines and outperformed all comparators ( r = 0.84, RMSE = 1.00). ROC and calibration analyses confirmed well-calibrated posterior probabilities. Feature-weight posteriors were enriched (p < 10⁻¹⁰) for known QTL and successfully rediscovered the DGAT1 milk-fat locus. Profiling showed sustained 80–95 % GPU utilisation with stable memory footprints, validating computational feasibility. Hyper-parameter sweeps revealed model robustness to learning-rate and prior-variance choices. Conclusions: Marrying multi-omics databases with Bayesian deep learning yields state-of- the-art accuracy, interpretable biology, and scalable compute in livestock genomic prediction. The approach provides a blueprint for data-driven animal breeding and can be readily extended to other species and omics layers.