This program combines blupf90, remlf90, and airemlf90.
For a general understanding of the parameter file, check the manual available at documentation.
It has some new features such as the computation of accuracies of (G)EBV based on PEV, and the ability to save solutions with original ID.
See preGSf90 for options using genomic (SNP) information.
Hint: type blupf90+ --help to see all the blupf90+ options or blupf90+ --help-genomic to see all the genomic options blupf90+ can take.
OPTION conv_crit 1e-12
Set convergence criteria (default 1e-12).
OPTION maxrounds 10000
Set maximum number of rounds (default 5000).
OPTION solv_method FSPAK
Selection solutions by FSPAK, SOR or PCG (default PCG).
OPTION r_factor 1.6
Set relaxation factor for SOR (default 1.4).
OPTION sol se
Store solutions and standard errors.
OPTION store_pev_pec 6
Store triangular matrices of standard errors and its covariances for correlated random effects such as direct-maternal effects and random-regression effects in “pev_pec_bf90”.
OPTION missing -999
Specify missing observations in the data file. Must be an integer value (default is 0).
OPTION residual
y-hat and residual will be included in “yhat_residual”.
OPTION blksize 3
Set a block size for preconditioner (default 1) to accelerate convergence (usually 2 to 5 times faster). For a multi-trait model, use the number of traits. This option is extremely important to ensure convergence in multi-trait models.
OPTION prior_solutions
The previous solution file will be used to start the iteration. Additional software is required to set up files correctly before using this option.
OPTION stdresidual
y-hat and student residuals will be included in “yhat_student_residual”.
OPTION check_levels 0
Check levels (default 1 = true).
OPTION use_yams
Run the program with YAMS (modified FSPAK).
OPTION SNP_file snp
Specify the SNP file name to use genotype data.
OPTION snp_p_value
Computes the elements of the inverse of the Mixed Model Equations that are needed for exact GWAS with p-values using postGSf90. This requires quite a lot of memory and time. For details see https://doi.org/10.1101/555243.
blupf90+ allows storing solutions with original ID. To do it, simply add the following option:
OPTION origID
The output is trait effect level original_id solution, and is stored in solutions.original.
Reliabilities are 1-PEV/($\sigma\mathbf{}_{u}^{2}$(1 + F))
where PEV is the prediction error variance of a (G)EBV. Reliabilities are the squared correlation between (True) BV and (G)EBV. Accuracies are the square root of reliabilities and thus are the correlation(TBV,(G)EBV).
OPTION store_accuracy eff
Stores reliabilities (not accuracies in spite of the name) based on PEV, where eff is the number of the animal effect.
By default, it uses inbreeding (F) in the denominator of the reliability formula: reliability = 1-PEV/($\sigma\mathbf{}_{u}^{2}$(1 + F)) Aguilar et al. (2020).
It uses inbreeding based on $\mathbf{A}$ or $\mathbf{H}$ from the direct inversion of $\mathbf{A}^{-1}$ or $\mathbf{H}^{-1}$, whichever is being used.
The resulting file, acc_bf90, contains: trait, effect, level, solutions, reliabilities.
OPTION store_accuracy eff orig
Stores reliabilities based on PEV, where eff is the number of the animal effect, and orig should be used to output reliabilities with original ID. Combine with OPTION origID if solutions should also be output with original ID.
The resulting file, acc_bf90, contains: trait, effect, level, original_ID, solutions, reliabilities.
OPTION acctype 1.0
Select 1.0 for dairy cattle (Reliability) or 0.5 for beef cattle (BIF accuracy) (default 1.0).
OPTION correct_accuracy_by_inbreeding filaname
filename is the name of the inbreeding file if other than renf90.inb
OPTION correct_accuracy_by_inbreeding_direct 0
This option turns off the inbreeding correction in the reliability formula.
OPTION method VCE
Runs airemlf90 for variance component estimation (default running blupf90)
OPTION conv_crit 1d-12
Convergence criterion (default 1d-10).
OPTION maxrounds 1000
Maximum rounds (default 5000). When the number = 0, the program calculates BLUP without iterating REML and some statistics (-2logL, AIC, SE for (co)variances, …).
* OPTION EM-REML n (with n an integer): Runs first EM-REML either for n rounds or until convergence (whatever comes first). then switches to AI-REML until convergence for a maximum of maxrounds, at the end it shows AI-REML output. It may happen that EM-REML converges and then AI-REML diverge.
* OPTION EM-REML or OPTION EM-REML pure: Runs EM until convergence or maxrounds, showing EM output. This is equivalent to REMLF90. It does not provide Average Information matrix or standard error of the estimates.
* OPTION EM-REML AI or OPTION EM-REML ai: Runs EM until convergence and then calculates the Average Information matrix. It can provide standard error of the estimates.
* OPTION EM-REML AI conv or OPTION EM-REML ai conv: Runs EM-REML until convergence and then switches to AI-REML rounds until either AI-REML convergence or maxrounds. It also provides Average Information output and can show s.e. of the estimates. It may happen that EM-REML converges and then AI-REML diverge.
For very complex models of if there is little information, OPTION EM-REML or OPTION EM-REML AI are recommended.
OPTION use_yams
Runs the program with YAMS (modified FSPAK). The computing time can be dramatically improved.
OPTION set_eig 1d-12
Tolerance (or precision) (default 1d-18) for positive definite matrix and g-inverse subroutines.
Convergence may be much faster by changing this value.
OPTION sol se
Stores solutions and standard errors.
OPTION store_pev_pec 6
Stores triangular matrices of standard errors and its covariances for correlated random effects such as direct-maternal effects and random-regression effects in “pev_pec_bf90”.
OPTION residual
y-hat and residuals will be included in “yhat_residual”.
OPTION missing -999
Specifies the missing value (default 0) in integer.
OPTION constant_var 5 1 2 ...
5: effect number
1: first trait number
2: second trait number
implying the covariance between traits 1 and 2 for effect 5.
Heterogeneous residual variances for a single trait
OPTION hetres_pos 10 11
Specifies the position of covariables.
OPTION hetres_pol 4.0 0.1 0.1
Initial values of coefficients for heterogeneous residual variances using ln(a0, a1, a2, …) to make these values.
Heterogeneous residual variances for multiple traits
Convergence will be very slow with multiple trait heterogeneous residual variances
OPTION hetres_pos 10 10 11 11
or
OPTION hetres_pos 10 11 12 13
Specify the position of covariables (trait first).
“10 10” or “10 11” could be linear for first and second traits.
“11 11” or “12 13” could be quadratic.
OPTION hetres_pol 4.0 4.0 0.1 0.1 0.01 0.01
Initial values of coefficients for heterogeneous residual variances using ln(a0, a1, a2, …) to make these values (trait first).
“4.0 4.0” are intercept for first and second traits.
“0.1 0.1” could be linear and “0.01 0.01” could be quadratic.
To transform back to the original scale, use exp(a0+a1*X1+a2*X2).
OPTION SNP_file snp
Specify the SNP file name to use genotype data.
OPTION se_covar_function <label> <function>
As an alternative of SE, calculate SD for function of (co)variances by repeated sampling of parameters estimates from their asymptotic multivariate normal distribution, following ideas presented by Meyer and Houle 2013 (Meyer, Karin, and David Houle. “Sampling based approximation of confidence intervals for functions of genetic covariance matrices.” Proc. Assoc. Advmt. Anim. Breed. Genet. Vol. 20. 2013.) .
<label>
A name for a particular function (e.g., P1 for phenotypic variance of trait 1, H2_1 for heritability for trait 1, rg12 for genetic correlation between traits 1 and 2, …).
<function>
A formula to calculate a function of (co)variances to estimate SD. All terms of the function should be written with NO spaces.
Each term of the function corresponds to (co)variance elements and could include any random effects (G) and residual (R) (co)variances.
Notation is with reference to the effect number and the trait number (G_eff1_eff2_trt1_trt2) that indicate the element of the (co)variance matrix for random effect eff1 and eff2 and trt1 and trt2,
where eff1 and eff2 are effect numbers 1 and 2, and trt1 and trt2 are trait numbers 1 and 2.
R_trt1_trt1 indicates the element of the residual (co)variance matrix for traits 1 and 2.
Several functions could be added, with one OPTION line per function.
Examples:
OPTION se_covar_function P G_2_2_1_1+G_2_3_1_1+G_3_3_1_1+G_4_4_1_1+R_1_1
OPTION se_covar_function H2d G_2_2_1_1/(G_2_2_1_1+G_2_3_1_1+G_3_3_1_1+G_4_4_1_1+R_1_1)
OPTION se_covar_function H2t (G_2_2_1_1+1.5*G_2_3_1_1+0.5*G_3_3_1_1)/(G_2_2_1_1+G_2_3_1_1+G_3_3_1_1+G_4_4_1_1+R_1_1)
OPTION se_covar_function rg12 G_2_2_1_2/(G_2_2_1_1*G_2_2_2_2)**0.5
The first function calculates the SD for the total variance for a maternal model with permanent maternal effect, where 2 and 3 are the effect number for the direct and maternal additive genetic effects respectively, and 4 is the effect number for the maternal permanent random effect.
The second function calculates the heritability for the direct component.
The third function the total heritability.
The fourth function calculates the SD of the genetic correlation between traits 1 and 2 for the direct genetic effect (effect number 2)
OPTION samples_se_covar_function <n>
Set the number of samples to calculate SE for function of (co)variances.
default value 10000
OPTION out_se_covar_function
Indicate to store in file samples of (co)variances function for postprocessing (histogram, etc.)
When the covariance matrix is close to non-positive definite, BLUPF90+ with “OPTION method VCE” (AIREMLF90) may not converge.
There are two options you might want to try the following options.
1. Change the tolerance value (xx) in the option:
OPTION set_eig xx
to a very strict value (e.g., 1d-24) or a lenient value (e.g., 1d-04)
2. Use options described above to use EM-REML alone or before inside AI-REML.
3. When traits and/or correlated random effects are highly correlated (e.g., > 0.95), removing one correlated trait or effect is one option. Since two traits (effects) are highly correlated (especially genetically), there is no reason to select two traits together. This option may reduce the phenotyping cost.
4. When genetic covariances among multiple traits cannot be accurately estimated for some reason, a separate analysis using a single-trait model would be a better choice. It can happens, especially when one trait has too many missing observations and/or weak relationships among animals. The single-trait model will be much better than the multiple-trait model.
Also, some options (2 and 3) described below for BLUP could be useful for VCE.
When the convergence is very slow or never reached, there are several reasons.
1. The covariance matrix G and/or R is close to non-positive definite.
In this case, change the tolerance value (xx) in the option:
OPTION set_eig xx
to a lenient value (e.g., 1d-04),
which is similar to re-parameterization for G = VDV' using larger eigenvalue(s) in D.
2. Remove the old data, which the contribution to the current active population is minimal.
3. Remove non-contribution animals, which have no relationships with other animals that have phenotypes.
4. Remove animals with no phenotypes in the input data file. If those animals have relationships in the pedigree and/or the genotypes, they will be included in the renaddXX.ped (output pedigree file) after running renumf90. So, they do not need to be included in the phenotypic data file when they have no phenotypes.
5. Re-estimate the variance-covariance matrices accurately using the current data.
The program accepts now different WEIGHTS for different traits. This works both for BLUP and for REML estimates. It has also been introduced in renumf90 and BLUP90IOD3 (new BLUP90IOD) but not in gibbsf90+.
With this modification weight_y becomes a ntrait x ntrait matrix. The way it works is that, in the MME, the weight applied to trait i is weight_y(i,i) and to the pairs trait i and trait j is weight_y(i,j) where in fact weight_y(i,j)=sqrt(weight_y(i,i)*weight_i(j,j)).
This corresponds to a model where the residual variance of record say k is Wk R Wk where Wk contains sqrt(1/weight(traits)) of record k in the diagonal so all is coherent with multivariate normality.
The syntax is:
WEIGHT(S) 7 8 9
with “old” syntax for a single weight across all traits, programs always work:
WEIGHT(S) 7
or using the new syntax:
WEIGHT(S) 7 7 7
WEIGHT(S)
the program stops when there are less weights than traits, e.g.:
NUMBER_OF_TRAITS
3
...
WEIGHT(S)
7 7
with the following message:
Values for WEIGHT can be 'empty line' (no weight), a single number (same weight for all lines) or 'ntrait' numbers (different weights for each trait)
OPTION omit_ainv
This option prohibits the program from creating $\mathbf{A}^{-1}$. It is especially useful for GBLUP. For example, if you would like to perform the exact GBLUP, you can put the following options to your parameter file.
OPTION omit_ainv OPTION TauOmega 1.0 0.0 OPTION AlphaBeta 0.95 0.05
With the above options, the program doesn't create $\mathbf{A}^{-1}$ but calculates $\tau\mathbf{G}^{-1}-\omega\mathbf{A}_{22}^{-1}$. When the omega ($\omega$) is zero, only $\mathbf{G}^{-1}$ will be included in the equations. $\mathbf{G}$ is blended with $\mathbf{A}_{22}$ as $\alpha\mathbf{G}+\beta\mathbf{A}_{22}$ before the inversion ($\alpha=0.95$ and $\beta=0.05$ in this case).
Assuming a single-trait ssGBLUP, the mixed model equations are as follows.
\( \left[ \begin{array}{ll} \mathbf{X}'\mathbf{X} & \mathbf{X}'\mathbf{Z}\\ \mathbf{Z}'\mathbf{X} & \mathbf{Z}'\mathbf{Z} + \lambda\mathbf{H}^{-1} \end{array} \right] \left[ \begin{array}{c} \mathbf{\hat{b}}\\ \mathbf{\hat{u}} \end{array} \right] = \left[ \begin{array}{c} \mathbf{X}'\mathbf{y}\\ \mathbf{Z}'\mathbf{y} \end{array} \right] \)
where $\mathbf{H}$ is a matrix combining additive genetic relationship matrices and a genomic relationship matrix.
\( \mathbf{H}^{-1} = \mathbf{A}^{-1} + \left[ \begin{array}{cc} \mathbf{0} & \mathbf{0} \\ \mathbf{0} & \tau\mathbf{G}^{-1}-\omega\mathbf{A}_{22}^{-1} \end{array} \right] \)
If we omit $\mathbf{A}^{-1}$ and $\mathbf{A}_{22}^{-1}$, the equations are equivalent to GBLUP. GBLUP by BLUPF90 was not so easy because the program creates $\mathbf{A}^{-1}$ by default and there was no way to avoid this behavior. The new option removes $\mathbf{A}^{-1}$ from the equations so GBLUP will be easily performed.