Accuracies using blupf90 suite including accf90GS

The blupf90 suite of programs has a few options to compute individual reliabilities.

The first and exact one is using blupf90 itself, by inversion of the Mixed Model Equations, using

OPTION sol se

This is however only doable for small-medium data sets (up to maybe 3 million equations).

The second one, which is more appropriate for threshold traits, is using Gibbs sampler by, for instance, gibbsf90test computing the posterior variance of the EBVs using

OPTION fixed_var all

For an example, see here. Still, this requires large memory and lots of patience.

The third one, which is needed in large data sets, is by approximation. This program, including now genomic information, is accf90GS and is only available under special agreement. (If you can use blupf90 and invert the MME, do it - it is much better than using accf90GS).  Several approximations exist, the ones that are used in this program combine essentially three papers:

• Misztal, I., & Wiggans, G. R. (1988). Approximation of prediction error variance in large-scale animal models. Journal of Dairy Science, 71, 27-32.
• Sanchez, J. P., Misztal, I., & Bertrand, J. K. (2008). Evaluation of methods for computing approximate accuracies of predicted breeding values in maternal random regression models for growth traits in beef cattle. Journal of animal science, 86(5), 1057-1066.
• Misztal, I., Tsuruta, S., Aguilar, I., Legarra, A., VanRaden, P. M., & Lawlor, T. J. (2013). Methods to approximate reliabilities in single-step genomic evaluation. Journal of dairy science, 96(1), 647-654.

Running accf90GS is rather tricky. In my case, it gets even worse because I do not use renumf90. So this is how I do:

First, I run preGSf90 to get the diagonal of the matrix $latex G^{-1}-A_{22}^{-1} $latex  using, in addition to the usual options

OPTION SNP_file ../typ.blupf90

OPTION CreateGimA22i

the specific option

OPTION saveDiagGimA22iRen

This option saves the elements of the diagonal of this matrix with the same numbers as the first column of the pedigree file, and not in the order 1, 2... as it is ordered internally.

Second, I run my SSGBLUP genetic evaluation as usual.

Third,  I run Accf90GS with options

OPTION DiagGinv_file ./DiagGimA22i_ren.txt

OPTION adj_zeta 1.0

OPTION conv_crit 1d-06

where  tells the program where to find $diag(G^{-1}-A_{22}^{-1})$ , adj_zeta 1.0 is a measure of the amount of information added by genomics (usually 1) and conv_crit 1d-06 tells the program to iterate but not too much.