Appendix
The additive relationship matrix-adjustment for self-fertilization
In plant breeding, the test lines that are included in trials are often the result of five or six generations of self-fertilization. The method of Henderson (1976) was developed for use in animal pedigrees, and as such requires for any particular line that is a result of n generations of self-fertilization that all the previous n − 1 generations of lines involved in its development are included in the pedigree. Clearly, in plant breeding trials where each test line has undergone the self-fertilization process up to n times, this would require an (unnecessarily) large pedigree to be recorded in order to obtain accurate estimates of a
jt
. A modification in the calculation of the inbreeding coefficient F
j
and therefore in the a
jj
value, can be incorporated into the algorithm, so that it is unnecessary to include the n − 1 generation of lines in the pedigree, just the number of generations n of self-fertilization need be recorded for each line.
If both parents, s and d of individual j are known then, the adjustment under n generations of self-fertilization is given by
$$a_{{jj}} = 2-0.5^{{n-1}} + 0.5^{n} a_{{sd}}$$
(9)
which reduces to Henderson’s equation under no self-fertilization, i.e. n = 1, also note that a
jj
tends to 2 as n tends to infinity.
Under n generations of self-fertilization, when one parent is known or when no parents are known the value of a
jj
can be shown to be
$$a_{{jj}} = 2-0.5^{{n - 1}}. $$
The coefficient of parentage matrix-adjustment for self-fertilization
The method of Sneller (1994) does not take into consideration self-fertilization. A modification in the calculation of the inbreeding coefficient F
j
and therefore f
jj
is necessary when dealing with individuals that have been self-fertilized for n generations.
Under self-fertilization, the coefficient of parentage f
jj
of j in the nth generation is given by half equation Eq. 9 as follows:
$$f_{{jj}} = 1-0.5^{n} + 0.5^{n} f_{{sd}}$$
(10)
When one parent is known or when no parents are known the value of f
jj
is f
jj
= 1 − 0.5n
ASReml code for fitting the Pedigree model 2
The following is the code for the .as ASReml file used for fitting the Pedigree model to a trial.
The stage3.giv is a file containing the inverse of the additive relationship matrix. ASReml requires a file which is just the lower triangle of this matrix. It is important to ensure that the numbering of lines in knownped factor corresponds directly to the ordering of rows and columns in the “.giv” file, so that row one and column one of the A inverse matrix contain the additive relationships of individual 1, which should correspondingly be labeled as 1 in the knownped factor. The “.giv” file can be created in ASReml if a pedigree file is supplied, and ASReml now implements the adjustment for inbred lines.
The stage3rba.asd is a text file containing the data. The knownped and filler columns have been created from the line column in which the lines are numbered from 1:253. In particular, the knownped is a column which has been defined as a factor with 129 levels. The levels correspond to the lines with known pedigree. It has “NA”s in the positions which correspond to filler lines. The filler is a column which has been defined as a factor with 124 levels, filler lines are defined as 1:124 and lines which have pedigree have “NA”s. The ped column is a factor which has two levels so that separate overall means can be fitted for filler lines and lines with known pedigree.
The additive genetic effect for each line is fitted by including the term knownped in the random part of the model specification and the epistatic genetic effect for each line is fitted by including the term ide(knownped) in the random part of the model specification, the units term is the measurement error term.
The last two lines are the predict statements to obtain the elements of the estimated prediction error variance matrix, so that the generalized heritability can be calculated. ASReml places these in the “.pvs” file. The estimated prediction error variance of the total genetic effects is used for calculating a broad sense heritability and those of the additive effects for calculating a narrow sense heritability. Calculation of generalized heritability was carried out using R (R Development Core Team 2005). The R code is available from the corresponding author.
Generalized definition of heritability
The Lagrangian given by Eq. 6 is to be optimized with respect to c. Thus, differentiating
\(\L_{{{\mathbf{c}}}} \) with respect to c and setting to zero, we find
$${\mathbf{Z}}^{T}_{g} {\mathbf{P}}_{v} {\mathbf{Z}}_{g} {\mathbf{Gc}} = \lambda {\mathbf{c}}.$$
(11)
Thus, c is an eigenvector of the matrix Z
T
g
P
v
Z
g
G with eigenvalue λ. Not only can the c that maximizes the squared correlation be found, but a complete set of eigenvectors c for Z
T
g
P
v
Z
g
G with associated eigenvalues. Notice that from Eq. 11
$$\begin{aligned} {\mathbf{c}}^{T} {\mathbf{GZ}}^{T}_{g} {\mathbf{P}}_{v} {\mathbf{Z}}_{g} {\mathbf{Gc}}&=\lambda {\mathbf{c}}^{T} {\mathbf{Gc}}\\ &= \lambda \\ \end{aligned} $$
using the constraint. Thus, the eigenvalues provide a set of heritability components that can be used to provide an overall measure of heritability.
From results on mixed models, GZ
T
g
P
v
Z
g
G = G − (Z
T
g
SZ
g
+ G
−1)−1 where S = R
−1 − R
−1
X(X
T
R
−1
X)−1
X
T
R
−1. Now (Z
T
g
SZ
g
+ G
−1)−1 = C
ZZ is the partition of the inverse of the mixed model coefficient matrix corresponding to g. This latter term C
ZZ is also equivalent to the prediction error variance matrix (i.e.
\(\hbox{var} (\tilde{\mathbf{g}} - {\mathbf{g}})\)), an estimate of which is available in the software ASReml (Gilmour et al. 2005) via the predict statement. So
$${\mathbf{Z}}^{T}_{g} {\mathbf{P}}_{v} {\mathbf{Z}}_{g} {\mathbf{G}} = {\mathbf{I}}_{m} - {\mathbf{G}}^{{- 1}} {\mathbf{C}}^{{ZZ}} $$
(12)
and eigenvalues of this matrix are required to determine the generalized heritability.