synbreed (version 0.12-9)

gpMod: Genomic predictions models for objects of class gpData

Description

This function fits genomic prediction models based on phenotypic and genotypic data in an object of class gpData. The possible models are Best Linear Unbiased Prediction (BLUP) using a pedigree-based or a marker-based genetic relationship matrix and Bayesian Lasso (BL) or Bayesian Ridge regression (BRR). BLUP models are fitted using the REML implementation of the regress package (Clifford and McCullagh, 2012). The Bayesian regression models are fitted using the Gibbs-Sampler of the BGLR package (de los Campos and Perez, 2010). The covariance structure in the BLUP model is defined by an object of class relationshipMatrix. The training set for the model fit consists of all individuals with phenotypes and genotypes. All data is restricted to individuals from the training set used to fit the model.

Usage

gpMod(gpData, model=c("BLUP","BL","BRR"), kin=NULL, predict=FALSE, trait=1,
      repl=NULL, markerEffects=FALSE, fixed=NULL, random=NULL, ...)

Arguments

gpData

object of class gpData

model

character. Type of genomic prediction model. "BLUP" indicates best linear unbiased prediction (BLUP) using REML for both pedigree-based (P-BLUP) and marker-based (G-BLUP) model. "BL" and "BRR" indicate Bayesian Lasso and Bayesian Ridge Regression, respectively.

kin

object of class relationshipMatrix (only required for model = "BLUP"). Use a pedigree-based kinship to evaluate P-BLUP or a marker-based kinship to evaluate G-BLUP. For "BL" and "BRR", also a kinship structure may be used as additional polygenic effect \(u\) in the Bayesian regression models (see BGLR package).

predict

logical. If TRUE, genetic values will be predicted for genotyped but not phenotyped individuals. Default is FALSE. Note that this option is only meaningful for marker-based models. For pedigree-based model, please use function predict.gpMod.

trait

numeric or character. A vector with names or numbers of the traits to fit the model

repl

numeric or character. A vector with names or numbers of the repeated values of gpData$pheno to fit the model

markerEffects

logical. Should marker effects be estimated for a G-BLUP model, i.e. RR-BLUP? In this case, argument kin is ignored (see Details). Plose note, that in this case also the variance components pertaining to model G-BLUP are reported instead of those from the G-BLUP model (see vignette). If the variance components are committed to crossVal, it must be guaranteed that there also the RR-BLUP model is used, e.g. no cov.matrix object should be specified.

fixed

A formula for fixed effects. The details of model specification are the same as for lm (only right hand side required). Only for model="BLUP".

random

A formula for random effects of the model. Specifies the matrices to include in the covariance structure. Each term is either a symmetric matrix, or a factor. Independent Gaussian random effects are included by passing the corresponding block factor. For mor details see regress. Only for model="BLUP"

…

further arguments to be used by the genomic prediction models, i.e. prior values and MCMC options for the BLR function (see BLR) or parameters for the REML algorithm in regress.

Value

Object of class gpMod which is a list of

fit

The model fit returned by the genomic prediction method

model

The model type, see 'Arguments'

y

The phenotypic records for the individuals in the training set

g

The predicted genetic values for the individuals in the training set

m

Predicted SNP effects (if available)

kin

Matrix kin

Details

By default, an overall mean is added to the model. If no kin is specified and model = "BLUP", a G-BLUP model will be fitted. For BLUP, further fixed and random effects can be added through the arguments fixed and random.

The marker effects \(\hat{m}\) in the RR-BLUP model (available with markerEffects) are calculated as $$\hat{m}= X'G^{-1}\hat{g}$$ with \(X\) being the marker matrix, \(G=XX'\) and \(hat{g}\) the vector of predicted genetic values.

Only a subset of the individuals - the training set - is used to fit the model. This contains all individuals with phenotypes and genotypes. If kin does not match the dimension of the training set (if, e.g. ancestors are included), the respective rows and columns from the trainings set are choosen.

References

Clifford D, McCullagh P (2012). regress: Gaussian Linear Models with Linear Covariance Structure. R package version 1.3-8, URL http://www.csiro.au.

Gustavo de los Campos and Paulino Perez Rodriguez, (2010). BLR: Bayesian Linear Regression. R package version 1.2. http://CRAN.R-project.org/package=BGLR

See Also

kin, crossVal

Examples

# NOT RUN {
library(synbreedData)
data(maize)
maizeC <- codeGeno(maize)

# pedigree-based (expected) kinship matrix
K <- kin(maizeC,ret="kin",DH=maize$covar$DH)

# marker-based (realized) relationship matrix
# divide by an additional factor 2
# because for testcross prediction the kinship of DH lines is used
U <- kin(maizeC,ret="realized")/2
# BLUP models
# P-BLUP
mod1 <- gpMod(maizeC,model="BLUP",kin=K)
# G-BLUP
mod2 <- gpMod(maizeC,model="BLUP",kin=U)

# Bayesian Lasso
prior <- list(varE=list(df=3,S=35),lambda = list(shape=0.52,rate=1e-4,value=20,type='random'))
mod3 <- gpMod(maizeC,model="BL",prior=prior,nIter=6000,burnIn=1000,thin=5)

summary(mod1)
summary(mod2)
summary(mod3)
# }