Uncertainty-aware breeding decisions: MCMC-based optimum contribution selection increases breeding decision robustness

Abstract Current optimum contribution selection (OCS) implementations use point estimates of estimated breeding values (EBVs), potentially leading to suboptimal selections when individuals have uncertain genetic evaluations. We developed a framework assessing how EBV uncertainty affects OCS decisions through MCMC-based approaches using the COSMO optimizer in Julia, evaluated on Norway spruce (Picea abies, n=5,525) and Loblolly pine (Pinus taeda, n=926) populations. Agreement between point estimate (MAP-OCS) and MCMC-OCS was surprisingly low: mean overlap of only 26.6 (4.8) individuals in Norway spruce genotyped subpopulation and 14.1 (3.6) in full pedigree, with Loblolly pine intermediate at 16.0 (9.6). Despite this low individual-level agreement, selection frequency across MCMC iterations corresponded well with EBV rankings (Spearman ρ = 0.782 for Norway spruce), confirming that higher-EBV individuals were preferentially selected under posterior uncertainty. To comprehensively quantify uncertainty impacts, we employed two complementary metrics: individual robustness scores measuring genetic gain stability upon candidate removal, and population-level contribution distribution metrics capturing concentration of genetic gain across selected individuals. Applying these metrics identified 25 high-risk individuals in Norway spruce and nine in Loblolly pine, and constrained exclusion of these individuals improved individual robustness by 16.5% in Loblolly pine (3.00% genetic gain loss) and 29.8% in Norway spruce (2.14% genetic gain loss). Our uncertainty-aware OCS framework successfully identifies unstable selections that may compromise long-term genetic gain, and we recommend assessing EBV uncertainty through posterior distributions and evaluating population-specific trade-offs when implementing uncertainty-aware selection strategies.