Multi-trait ensemble genomic prediction and simulations of recurrent selection highlight importance of complex trait genetic architecture in long-term genetic gains in wheat

Abstract

Crop breeders have achieved considerable genetic gain in genetically complex traits, such as grain yield, while maintaining genetic diversity. However, a focus on selection for yield has negatively impacted other important traits. To better understand this selection process, and how it might be optimised, we analysed extensive genotypic and phenotypic data from a diverse, 16-founder wheat multi-parent advanced generation inter-cross (MAGIC) population. Compared to single-trait models, multi-trait ensemble models increased prediction accuracy for almost 90% of traits, improving grain yield prediction accuracy by 3-52%. For complex traits, non-parametric models (Random Forest) also outperformed simplified, additive models (LASSO), increasing grain yield prediction accuracy by 10-36%. Simulations of recurrent genomic selection then showed that sustained greater forward prediction accuracy optimised long-term genetic gains. Simulations of selection on grain yield found indirect responses in related traits which involved optimisation of antagonistic trait relationships. We found that multi-trait selection indices can be used to optimise undesirable relationships, such as the trade-off between grain yield and protein content, or combine traits of interest, such as yield and weed competitive ability. Simulations of phenotypic selection found that including Random Forest rather than LASSO genetic models, and multi-trait rather than single-trait models as the true genetic model accelerated and extended long-term genetic gain whilst maintaining genetic diversity. These results suggest important roles of pleiotropy and epistasis in the wider context of wheat breeding programmes and provide insights into mechanisms for continued genetic gain in a limited gene pool and optimisation of multiple traits for crop improvement.