Combining genomic prediction and multi-trait indices through stochastic simulations: do index type and deployment order affect genetic gain?

Genomic selection (GS) has transformed plant breeding by enabling early selection and potentially reducing cycle length, but how to integrate GS with classical multi-trait selection indices remains unclear. We used stochastic simulations to compare seven strategies combining Smith-Hazel (SH), Pesek-Baker (PB), and empirical (EMP) indices with or without GS in a rice breeding program targeting increased grain yield (GY), reduced chalkiness rate (CR), and stable plant height (PH). Trait importance was set to 70% for GY, 15% for CR, and 15% for PH. After a phenotypic-selection burn-in phase, we simulated ten recurrent cycles and evaluated strategies using population mean, selection accuracy, and additive genetic variance over the years. We also considered scenarios with different breeding cycle lengths (GS-based: 3 years; Traditional: 5 years), but the same selection intensity per stage. Thus, PB provides the best balance in gain responses, in terms of both magnitude and direction, with one of the highest GY, the greatest reduction in CR, and PH stabilization. Accuracy and additive genetic losses were more pronounced when the phenotypically constructed index was used as a single target for genomic prediction, especially for CR and PH. Overall, trait-specific genomic prediction followed by index construction better preserved genetic information and produced responses more consistent with the multi-trait breeding objectives.