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

Abstract Optimum contribution selection (OCS) balances genetic gain and inbreeding by optimizing parental contributions to the next generation, but current implementations rely on point estimates of breeding values that discard the uncertainty inherent in genetic evaluations. We introduce CVaR-OCS, a novel formulation that incorporates the full posterior distribution of estimated breeding values (EBVs) directly into the OCS objective via Conditional Value at Risk (CVaR) (a coherent risk measure from financial portfolio theory) allowing a single optimization to simultaneously maximize expected genetic gain and protect against worst-case outcomes driven by EBV uncertainty. We evaluate CVaR-OCS on a simulated multi-generation genomic selection dataset with known true breeding values (QTL-MAS 2010; n=3,226), enabling direct comparison to an oracle solution, and on Norway spruce (Picea abies  n=5,525) and Loblolly pine (Pinus taeda  n=926) forest tree breeding progeny trials, with consistent tail-gain protection observed across all three datasets. On the simulated dataset, point-estimate MAP-OCS recovered only 78% of the genetic gain achieved by an oracle solution with access to true breeding values, illustrating the cost of ignoring predictive uncertainty; CVaR-OCS attained comparable expected gain while improving tail-gain security (CVaR95+0.69%) and recovering an additional oracle-optimal individual. In Norway spruce, the recommended CVaR-OCS operating point improved tail-gain security by 6.60% and broadened the selection base from 145 to 159 individuals at a genetic gain cost of only 0.70%. Complementary MCMC-based robustness scores revealed that 25 MAP-OCS selections in Norway spruce were unstable across the posterior distribution; post-hoc exclusion of these individuals failed to improve tail-gain security, motivating the principled CVaR-OCS approach. CVaR-OCS provides breeders with a principled, computationally efficient tool for uncertainty-aware selection decisions, and multi-generation simulation studies are needed to fully characterize its long-term effects on genetic gain trajectories and inbreeding accumulation.