Simposio: Application of Mixed Models in Plant Breeding with New Biotechnologies

MÓNICA BALZARINI

Florianópolis, Brasil.

Congreso; XXVth International Biometric Conference (IBC-Floripa-2010); 2010

Mixed modeling has become a major area of biometrical research including linear, non-linear and generalized mixed models. Mixed models are statistical models with fixed and random effects that can be applied in situations where multiple correlated measurements are made on each unit of interest; they handle data where observations are not independent. The statistical model is specified in a hierarchical fashion assuming some parameters as random; the unknown parameters are estimated from the data. Increasing work on computation of maximum likelihood and Bayesian estimates of mixed models facilitates the practical approximation of best linear unbiased estimates (BLUE) of fixed effects and best linear unbiased predictions (BLUP) of random effects for complicated models and large datasets. The mixed effects model treats correlated data adequately and assumes different sources of variation: within and between clusters of data. Two types of coefficients are distinguished in the mixed model: population-averaged, allowing the same type of inference as in classical statistics from adjusted means and cluster (or subject)?specific inference for conditioning it to some specific random effects. Correlated trait values between relatives and between genotypic responses across some environments made mixed models a crucial tool for data analysis in plant breeding studies involving phenotypic data of related genotypes evaluated in a series of replicated field trials grown across multiple locations and years. Mixed models are able to incorporate pedigree information, to combine information across multiple environments and to cope with the common unbalancedness in plant breeding where the set of evaluated genotypes often changes between years and also between trials within year. Mixed models produce reliable inference through the explicit modeling of correlations induced by genetic and environmental causes. Currently, marker-assisted selection has become another standard tool in breeding programs and the selection of genetically superior individuals using molecular marker alleles linked to quantitative trait loci (QTL) affecting trait variation is common. The mixed-model framework is also powerful for this type of association studies in plant species. Their application can lead to substantially different conclusions in genotype selection within a breeding program compared to conventional analysis assuming independence. Including pedigree information or molecular information creates a connection network among data that improves estimates of individual genotype performances, genetic associations and genomic effects of trait variation. Additionally, partitioning of variance into within- and between- cluster components provides meaningful results for breeders such as estimations of additive and non-additive genetic effects, a means of predicting breeding values, and the possibility to explore genotype-by-environment interactions as well as differences in the effect of genomic regions across environments. In this context, we discuss about the formulation, estimation and goodness of fit of differing mixed models correcting for genetic relatedness and experimental data structure used in classical and molecular plant breeding.