Skip to contents

RR-BLUP Model 2

Source: R/GS.R
RRBLUP2.Rd

Fits an RR-BLUP model for genomic predictions. This implementation is meant for situations where RRBLUP is too slow. Note that RRBLUP2 is only faster in certain situations, see details below. Most users should use RRBLUP.

Usage

RRBLUP2(
  pop,
  traits = 1,
  use = "pheno",
  snpChip = 1,
  useQtl = FALSE,
  maxIter = 10,
  Vu = NULL,
  Ve = NULL,
  useEM = TRUE,
  tol = 1e-06,
  simParam = NULL,
  ...
)

Arguments

pop

a Pop-class to serve as the training population

traits

an integer indicating the trait to model, a trait name, or a function of the traits returning a single value. Unlike RRBLUP, only univariate models are supported.

use

train model using phenotypes "pheno", genetic values "gv", estimated breeding values "ebv", breeding values "bv", or randomly "rand"

snpChip

an integer indicating which SNP chip genotype to use

useQtl

should QTL genotypes be used instead of a SNP chip. If TRUE, snpChip specifies which trait's QTL to use, and thus these QTL may not match the QTL underlying the phenotype supplied in traits.

maxIter

maximum number of iterations.

Vu

marker effect variance. If value is NULL, a reasonable starting point is chosen automatically.

Ve

error variance. If value is NULL, a reasonable starting point is chosen automatically.

useEM

use EM to solve variance components. If false, the initial values are considered true.

tol

tolerance for EM algorithm convergence

simParam

an object of SimParam

...

additional arguments if using a function for traits

Details

The RRBLUP2 function works best when the number of markers is not too large. This is because it solves the RR-BLUP problem by setting up and solving Henderson's mixed model equations. Solving these equations involves a square matrix with dimensions equal to the number of fixed effects plus the number of random effects (markers). Whereas the RRBLUP function solves the RR-BLUP problem using the EMMA approach. This approach involves a square matrix with dimensions equal to the number of phenotypic records. This means that the RRBLUP2 function uses less memory than RRBLUP when the number of markers is approximately equal to or smaller than the number of phenotypic records.

The RRBLUP2 function is not recommend for cases where the variance components are unknown. This is uses the EM algorithm to solve for unknown variance components, which is generally considerably slower than the EMMA approach of RRBLUP. The number of iterations for the EM algorithm is set by maxIter. The default value is typically too small for convergence. When the algorithm fails to converge a warning is displayed, but results are given for the last iteration. These results may be "good enough". However we make no claim to this effect, because we can not generalize to all possible use cases.

The RRBLUP2 function can quickly solve the mixed model equations without estimating variance components. The variance components are set by defining Vu and Ve. Estimation of components is suppressed by setting useEM to false. This may be useful if the model is being retrained multiple times during the simulation. You could run RRBLUP function the first time the model is trained, and then use the variance components from this output for all future runs with the RRBLUP2 functions. Again, we can make no claim to the general robustness of this approach.

Examples

#Create founder haplotypes
founderPop = quickHaplo(nInd=10, nChr=1, segSites=20)

#Set simulation parameters
SP = SimParam$new(founderPop)
SP$addTraitA(10)
SP$setVarE(h2=0.5)
SP$addSnpChip(10)

#Create population
pop = newPop(founderPop, simParam=SP)

#Run GS model and set EBV
ans = RRBLUP2(pop, simParam=SP)
pop = setEBV(pop, ans, simParam=SP)

#Evaluate accuracy
cor(gv(pop), ebv(pop))
#>        est_GV_Trait1
#> Trait1      0.646779