Bayesian neural networks for genomic prediction: uncertainty quantification and SNP interpretation with SHAP and GWAS.

key messageThis study presents a Bayesian neural networks framework with LASSO regularization and the GSMeSP interpretability tool, enabling accurate, uncertainty-aware, and biologically interpretable genomic prediction. Deep learning offers significant potential for genomic prediction by modeling complex, nonlinear genotype-phenotype relationships. However, its application in plant breeding has been constrained by limited model interpretability and a lack of uncertainty quantification. To address these challenges, we developed a Bayesian neural networks (BNNs) framework incorporating least absolute shrinkage and selection operator (LASSO) regularization for multi-trait genomic prediction with credible uncertainty estimation. In parallel, we introduce GSMeSP, a novel interpretability framework that integrates SHapley Additive exPlanations (SHAP) with genome-wide association study (GWAS) signals to prioritize trait-associated single nucleotide polymorphisms (SNPs) from both statistical and biological perspectives. We applied this framework to a diverse panel of 1385 upland cotton (Gossypium hirsutum) accessions genotyped with over 12,000 SNPs, evaluating performance across multiple fiber-related traits. The BNNs model consistently outperformed conventional and deep learning benchmarks, achieving 0.46-47.85% improvements in predictive accuracy. Moreover, it generated trait- and sample-specific 95% credible intervals, enabling robust uncertainty quantification and more informed selection decisions. Using GSMeSP, we identified biologically meaningful loci, with a substantial proportion of top-ranked SNPs located in the D-subgenome. Notably, chromosome D05 emerged as a genomic hotspot enriched for SNPs associated with fiber length, lint percentage, and uniformity. By integrating high predictive performance, credible uncertainty estimation, and biologically grounded interpretability, our framework provides a transparent and robust deep learning approach to accelerate genomic selection in crop breeding programs.