Bayesian Multi-Trait Methods for Genetic Experiments with Complex Replication.

Genome-enabled breeding value estimation and prediction in plants serve as important starting points for breeding programs. To maximize their potential, it is necessary to evaluate phenotypes that are relevant in target field environments. Thus, complex experiments with a large environmental influence are often required. However, these experiments produce noisy data that may obscure the contribution of genetic variation. Most modern statistical methods to estimate and especially predict genetic parameters have been developed for animals, where each genetically distinct individual is evaluated only once and the confounding effects of population structure are not strong. These methods are not adequate for most applications in rice and other crop species. We have developed a general Bayesian multi-trait approach that accounts for noise in the data, including low heritability, unbalanced designs, missingness, outlier observations, and systematic experimental bias. We also implemented steps to correct for deep population structure (e.g., differentiation between O. s. indica and japonica). These approaches open the door to novel experimental designs. We illustrate our methods on simulated data and examples from experiments in rice. A C++ library, MuGen, that efficiently implements our techniques is freely available on GitHub (https://github.com/tonymugen/MuGen).