Theor Appl Genet (2016) 129:273–287
DOI 10.1007/s00122-015-2626-6
ORIGINAL ARTICLE
Multiple‑trait‑ and selection indices‑genomic predictions for grain
yield and protein content in rye for feeding purposes
Albert Wilhelm Schulthess1 · Yu Wang1,2 · Thomas Miedaner2 · Peer Wilde3 ·
Jochen C. Reif1 · Yusheng Zhao1
Received: 1 June 2015 / Accepted: 17 October 2015 / Published online: 3 November 2015
© Springer-Verlag Berlin Heidelberg 2015
Abstract
Key message Exploiting the benefits from multiple-
trait genomic selection for protein content prediction
relying on additional grain yield information within
training sets is a realistic genomic selection approach in
rye breeding.
Abstract Multiple-trait genomic selection (MTGS) was
specially designed to benefit from the information of
genetically correlated indicator traits in order to improve
genomic prediction accuracies. Two segregating F3:4 rye
testcross populations genotyped using diversity array tech-
nology markers and evaluated for grain yield (GY) and
protein content (PC) were considered. The aims of our
study were to explore the benefits of MTGS over single-
trait genomic selection (STGS) for GY and PC prediction
and to apply GS to predict different selection indices (SIs)
for GY and PC improvement. Our results using a two-trait
Communicated by J. Wang.
A. W. Schulthess and Y. Wang equally contributed to this work.
Electronic supplementary material The online version of this
article (doi:10.1007/s00122-015-2626-6) contains supplementary
material, which is available to authorized users.
* Yusheng Zhao
zhao@ipk‑gatersleben.de
Albert Wilhelm Schulthess
schulthess@ipk‑gatersleben.de
Yu Wang
wangy@ipk‑gatersleben.de
Thomas Miedaner
thomas.miedaner@uni‑hohenheim.de
Peer Wilde
peer.wilde@kws‑lochow.de
model (2TGS) empirically confirm that the ideal scenario
to exploit the benefits of MTGS would be when the predic-
tions of a relatively low heritable target trait with scarce
phenotypic records are supported by an intensively phe-
notyped genetically correlated indicator trait which has
higher heritability. This ideal scenario is expected for PC
in practice. According to our GS implementation, MTGS
can be performed in order to achieve more cycles of selec-
tion by unit of time. If the aim is to exclusively improve
the prediction accuracy of a scarcely phenotyped trait,
2TGS will be a more accurate approach than a three-trait
model which incorporates an additional correlated indica-
tor trait. In general for balanced phenotypic information,
we recommend to perform GS considering SIs as single
traits, this method being a simple, direct and efficient way
of prediction.
Abbreviations
2T Two-trait
3T Three-trait
2TGS Two-trait genomic selection
3TGS Three-trait genomic selection
BLUE(s) Best linear unbiased estimator(s)
Jochen C. Reif
reif@ipk‑gatersleben.de
1
Department of Breeding Research, Leibniz Institute of Plant
Genetics and Crop Plant Research (IPK), 06466 Gatersleben,
Germany
2
State Plant Breeding Institute, University of Hohenheim,
70593 Stuttgart, Germany
3
KWS LOCHOW GMBH, 29296 Bergen, Germany
13
274
BLUP(s) Best linear unbiased predictor(s)
CMS Cytoplasmic-male sterile
DArT Diversity array technology
GEBV(s) Genomic predicted breeding value(s)
GS Genomic selection
GY Grain yield
MT Multiple-trait
MTGS Multiple-trait genomic selection
NIRS Near-infrared reflectance spectroscopy
O-SI Smith–Hazel or optimum selection index
PC Protein content
QTL Quantitative trait loci
REML Restricted maximum likelihood
RD Rogers’ distance
R-SI Restricted selection index
SDS Sudden death syndrome
SEW Single ear weight
SI Selection index
ST Single-trait
STGS Single-trait genomic selection
Introduction
Genomic selection (GS) is specifically designed to predict
complex traits such as grain yield (GY). In GS, training
populations which are phenotyped and genotyped are used
to estimate the effects of genome-wide marker panels. The
marker effects are then applied to predict the trait perfor-
mance of genotyped but non-phenotyped selection candi-
dates (Meuwissen et al. 2001). GS has been successfully
applied in plant breeding to predict for instance the per-
formance of testcross populations of maize (Albrecht et al.
2011; Windhausen et al. 2012; Zhao et al. 2012), sugar beet
(Hofheinz et al. 2012), and rye (Wang et al. 2014) or the
line per se performance of diversity panels in wheat (Crossa
et al. 2010; Heffner et al. 2011; Rutkoski et al. 2012), rice
(Spindel et al. 2015), soybean (Jarquín et al. 2014; Shu
et al. 2013), and barley (Heslot et al. 2012). The major ben-
efit of GS relies on its potential of speeding up selection
cycles and, thus, enhancing the selection gains per unit of
time and cost (Endelman et al. 2013; Longin et al. 2015;
Rutkoski et al. 2012).
In practice, when selection is performed for the improve-
ment of the economic value of animals or plants, it is gen-
erally applied to several traits simultaneously (Falconer and
MacKay 1996). However, current studies in plant sciences
generally target single-trait (ST) GS approaches without
exploiting information from correlated traits. A phenotypic
correlation occurs when the phenotypic values for two traits
are associated due to genetic and non-genetic causes, plei-
otropy and linkage being the two main known reasons for a
nonzero genetic correlation (Bernardo 2010). Multiple-trait
13
Theor Appl Genet (2016) 129:273–287
genomic selection (MTGS) was originally designed to
profit from the information contained in correlated indi-
cator traits (Calus and Veerkamp 2011). Since then, this
approach has been relatively widely applied to simulated
(Calus and Veerkamp 2011; Guo et al. 2014; Hayashi and
Iwata 2013; Jia and Jannink 2012) and animal breeding
datasets (Christensen et al. 2012; Liu et al. 2014; Tsuruta
et al. 2011). Simulation studies have extensively examined
the potential benefits of MTGS, and their findings agree
with advantages of indirect selection (Falconer and Mac-
Kay 1996): prediction accuracies for a target trait with rela-
tively low heritability can be substantially improved when a
genetically correlated indicator trait with higher heritability
is considered within the genomic prediction models (Guo
et al. 2014; Hayashi and Iwata 2013; Jia and Jannink 2012).
In contrast, studies applying MTGS into real/experimental
plant breeding data which aimed to take advantage of cor-
related indicator traits are less frequent (Bao et al. 2015;
Jia and Jannink 2012; Rutkoski et al. 2012). These applied
GS studies have concluded that the benefits of MTGS over
STGS are not as substantial as in simulation studies when
non-phenotyped selection candidates are intended to be
predicted. In addition, simulation and applied MTGS stud-
ies have been mainly focused on the factors influencing the
potential superiority in terms of prediction accuracy that
these methods can achieve compared to STGS, but they
have ignored the implications of selection when the corre-
lations between traits are in practice undesirable for plant
breeders. Presumably, all these practical limitations have
hindered the routine implementation of MTGS by public
plant breeding institutions and companies so far.
Secondary traits (like grain protein content, PC) nega-
tively correlated to a primary trait (for example GY) can
be deteriorated by means of single-trait selection (Dolan
et al. 1996). In this sense, when plant breeders are trying to
improve several quantitative traits simultaneously, selection
using a total score or economic selection index (SI) is gen-
erally more efficient than selection based on independent
culling levels or tandem selection (Hazel and Lush 1942).
SIs are based on the overall performance of genotypes
and, thus, consider the ability of favorable levels for some
trait(s) to compensate for unfavorable levels in another
trait(s) (Dolan et al. 1996). The Smith–Hazel or optimum
SI (O-SI) (Hazel 1943; Smith 1936) was originally designed
to achieve the simultaneous improvement of several traits.
More than a decade later, Kempthorne and Nordskog
(1959) derived a restricted SI (R-SI) from the O-SI, which
is specially designed for situations when breeders want to
improve some traits but simultaneously holding other traits
at their average level. Crop studies have evaluated the appli-
cation of these indices to improve the overall genotype per-
formance according to different traits, including studies on
alfalfa (Elgin et al. 1970), oat (Dolan et al. 1996; Eagles and
Theor Appl Genet (2016) 129:273–287
Frey 1974), maize (Suwantaradon et al. 1975), and soybean
(Holbrook et al. 1989; Openshaw and Hadley 1984). In the
past years, SIs for multiple-trait (MT) selection have been
coupled with information of molecular markers underly-
ing quantitative trait loci (QTL) (Cerón-Rojas et al. 2008)
or also with high-density molecular marker data as in GS
(Bernardo 2010; Dekkers 2007). However, until now, SIs
coupled with state-of-the-art GS approaches have not been
extensively applied in plant studies.
In the European Union, rye (Secale cereale L.) pro-
duction is mostly concentrated in Poland and Germany.
The end-uses of rye in Germany comprise food purposes
(25 %), mainly bread making, feeding (32 %), mostly for
home-grown feed, and bioenergy production (37 %) (http://
www.praxisnah.de/index.cfm/article/7270.html). Besides
GY, that is important for all purposes, in the last decades
breeding favored goals for bread making, which would not
necessarily meet the market requirements for other end-
uses of rye (Miedaner et al. 2012). For instance, PC in rye
must be minimized for baking purposes; however, this par-
ticular trait should reach sufficient levels to be commer-
cially competitive with wheat when rye is used for animal
feeding. Since GY and PC are intended to be selected in
the same direction in rye for feeding purposes, a nega-
tive correlation between them would impose difficulties
to the simultaneous breeding of both traits. In this study,
two segregating F3:4 rye testcross populations were evalu-
ated for GY, PC, and single ear weight (SEW) in up to 15
environments. The main objectives of this work were: (1)
to explore the benefits of MTGS over STGS for the predic-
tion of GY and PC in rye using different sizes of balanced
and unbalanced training datasets concerning both traits
plus the SEW and (2) to combine SIs with GS approaches
for GY and PC improvement when rye is used for feeding
purposes.
Materials and methods
Plant material, field experiments and phenotypic data
analyses
The plant material used for this study is described in detail
by Miedaner et al. (2012). Briefly, two segregating popula-
tions were developed from the crosses Lo115-N × Lo90-N
(POP-A) and Lo115-N × Lo117-N (POP-B). Subsequently,
220 lines from each population were randomly selected
at the level of F3:4 generation and then crossed with an
unrelated cytoplasmic-male sterile (CMS) single cross
tester to obtain three-way hybrids (testcrosses) of the type
(A·B) × F3:4 (Hübner et al. 2013; Miedaner et al. 2012).
Originally, the 220 (A·B) × F3:4 hybrids plus the test-
crosses of the two parents of each population (each of them
275
repeated nine times) and additionally two common checks
were evaluated across 15 environments (see Table S1 for
field experiments details) under an incomplete 24 × 10 lat-
tice design with two replications (Miedaner et al. 2012).
From these experiments, data for GY (t/ha), PC (%) and
SEW (g) were obtained. From past studies, GY data were
available for all environments (Hübner et al. 2013; Mie-
daner et al. 2012; Wang et al. 2015), whereas PC and SEW
data were originally obtained for ten (Miedaner et al. 2012;
Wang et al. 2015) and six (Miedaner et al. 2012) environ-
ments, respectively. SEW data for six additional environ-
ments (denoted as BEK11n, PET10n, PET11n, WAL10n,
WAL11n and WOH10n in Table S1) are included here,
making a total of 12 environments for this trait. The PC
was determined by near-infrared reflectance spectroscopy
(NIRS) as outlined in detail by Miedaner et al. (2012) and
Wang et al. (2015).
Phenotypic analyses for each trait (GY, PC, and SEW)
were performed using a two-step approach as presented in
detail by Wang et al. (2014). Firstly, best linear unbiased
estimators (BLUEs) were calculated within environments
with a linear model which included an intercept and the
testcross progenies as fixed factors, whereas blocks, rep-
lications and the error component were assumed as ran-
dom. In a second step, BLUEs across environments were
computed considering a model with an overall mean and
the testcross progenies as fixed factors in addition to the
environments plus the genotype × environment interac-
tion of the testcross progenies and the residual effect as
random. BLUEs across environments for the testcross
progenies were used to compute the phenotypic correla-
tions among traits. In parallel, considering the models of
the first and second step but assuming testcross progenies
as random, the corresponding variance components were
estimated. Using these estimates, heritabilities (h2) were
σG2
computed as: h2 = σ2 σ2
, where σG2 , σG×E
2
G×E +
σG2 + Nr.Env. Eff.Error
Nr.Rep.×Nr.Env.
and 2
σEff.Error represent the genotypic, genotype × environ-
ment interaction of the testcross progenies and the effective
error variance components, respectively; whereas, Nr.Env.
and Nr.Rep. are the number of environments and replicates,
correspondingly (Piepho and Möhring 2007). Additionally,
using the BLUEs for the testcross progenies obtained from
the first step, additive genetic (G0) and residual (R0) q × q
variance–covariance matrices among the q traits were esti-
mated pre-specifying a genomic estimated kinship matrix
within an MT mixed-model constructed according to Hen-
derson and Quass (1976). This model included the general
and environmental means as fixed effects and the testcross
progenies and residuals as random factors assuming an
unstructured covariance matrix between traits. Denoting
the diagonal and off-diagonal elements of G0 as cii and cij,
respectively, the genetic correlations between traits i and
13
276 Theor Appl Genet (2016) 129:273–287

c
j were computed as: √ciiij×cjj . BLUEs and variance com- (m × q) × 1 vector of zeros and G = m−1 G0 ⊗ Im, Im
ponents estimations were carried out using the restricted being an m × m identity matrix, and ⊗ the direct product
maximum likelihood (REML) algorithm included in the operator between matrices (Searle 1982). Similarly, e is
ASReml-R package (Butler et al. 2009). assumed to have a multivariate normal distribution as
e ∼ N(0, R), where 0 is an N × 1 vector of zeros and R is a
Marker data and genetic analyses sub-matrix of R0 ⊗ In, corresponding In to an n × n iden-
tity matrix. The R matrix is obtained from the R0 ⊗ In
Using diversity array technology (DArT) markers, 201 and matrix by removing the rows and columns which corre-
219 lines were genotyped for POP-A and POP-B, respec- spond to the genotypes without records for each trait (Hen-
tively (Miedaner et al. 2012). After quality checks for these derson and Quass 1976). With the above assumptions, the
marker data, 394 and 584 DArT markers were retained for BLUEs of µ (µ ) and the best linear unbiased predictors
POP-A and POP-B, correspondingly (Wang et al. 2014). Sub- (BLUPs) for a ( a ) could be obtained by the following
sequently, DArT markers were used to calculate the Rogers’ mixed-model equations (Henderson 1975):
distance (RD) matrix for each population as stated by Reif
   −1  
et al. (2005), but in this case they were expressed as J-RD 
µ 1′N R−1 1N 1′N R−1 Z 1′N R−1 Y
= . (2)
to resemble a relationship matrix, J being an n × n matrix 
a Z R 1N Z R−1 Z + G−1
′ −1 ′ Z′ R−1 Y
whose every element are 1 (Searle 1982) and n corresponding
to the number of genotypes for the population. These matri- Genomic predicted breeding values (GEBVs) were then
ces were used as the pre-specified genomic estimated kinship computed as µ + ZU a, where ZU is the design matrix for the
matrices for the G0 and R0 matrices estimation. untested genotypes. In the present study, GY and PC were
jointly predicted using mainly a two-trait (2TGS) model
GS methods (q = 2) and secondly, by means of a three-trait (3TGS)
model (q = 3) including the SEW as a third correlated trait.
Based on Henderson and Quass (1976), ST and MTGS could
be presented as a single unified mixed-model. In general, if SIs for GY and PC and their genomic predictions
n different genotypes or individuals are genotyped for m loci
(markers) and have ni phenotypic records available for q dif- For the construction of an SI for GY and PC, the O-SI and
ferent traits (with ni ≤ n and i = 1, 2, . . . , q), then the uni- R-SI were employed. A general SI for the ith genotype is
fied mixed-model to predict marker effects can be defined as: defined as SI i = p′i b,, where p′i corresponds to a 1 × q row
vector with the phenotypic values for the traits coming from
Y = 1N µ + Za + e, (1) the ith genotype and b is a q × 1 vector containing the index
where weights for each of the q traits (Lin and Allaire 1977). Then,
      for the O-SI (Hazel 1943; Smith 1936), b = P−1 G0 w, with
Y1 µ1 1n1 0 0
 Y2   µ2   0 1 ···
0  P being a q × q phenotypic variance–covariance matrix
     n2 
Y=
 .. ,
 µ=
 .. ,
 1N = 
 .. .. . ,
 between traits, G0 coming from the phenotypic data analy-
 .   .   . . ..  ses and w being a q × 1 vector which contains the rela-
Yq µq 0 0 · · · 1 nq tive economic weights for the q traits under selection. The
R0
      P matrix was derived as P = G0 + Nr.Env. , with G0 and
Z1 0 0 a1 e1
 0 Z     R0 estimated from the phenotypic data analyses. For the
0 
· · ·
 2   a2   e2 
Z= . 
. . .. , a= .. , and e =  ..  . R-SI, index weights are calculated by the following equa-
 −1
 .    
 . . .   .   .  tion: b = Iq − P Gs (Gs P Gs ) Gs P G0 w, with P
−1 ′ −1 ′ −1
0 0 · · · Zq aq eq and G0 as previously defined for the O-SI, Iq representing a
For trait i (i = 1, 2, …, q), Y i is the ni × 1 vector of pheno- q × q identity matrix and Gs corresponding to a G0 matrix
typic values, μi corresponds to the overall mean, Zi repre- whose q – s rows were deleted, s being the number of traits
sents the ni × m design matrix of marker effects, ai refers whose genetic gains should be set to zero (Bernardo 2010;
to the m × 1 vector of marker effects, 1ni denotes an ni × 1 Kempthorne and Nordskog 1959). Using an R-SI for GY
vector whose every element is a unit (Searle 1982), and ei and PC, the two corresponding scenarios, when breeders
is the ni × 1 vector of residual variation. Thetotal number want to select for GY (R-SI_GY) or PC (R-SI_PC) main-
q
of phenotypic records corresponds to N = 1 ni. In this taining the remaining trait at the mean level, were taken
sense, it can be deduced that an ST model corresponds to a on count. When an R-SI is calculated based on two traits,
particular case of the unified mixed-model, when q = 1. In the economic weights become irrelevant (Kempthorne and
parallel, it is assumed that a follows a multivariate normal Nordskog 1959). Hence, only a 1:1 price ratio between GY
distribution in the way: a ∼ N(0, G), where 0 is an and PC was considered for its calculation. However, for the

13
Theor Appl Genet (2016) 129:273–287
O-SI, the economic weights play an important role; hence
an approximate 10:3 price ratio between GY and PC was
additionally considered for this SI. This last price ratio
was assumed based on the Canadian Wheat Board (CWB)
prices and PC bonifications for small grain cereals (http://
www.cwb.ca/pricing) and the GY means for POP-A and
POP-B, respectively.
Based on suggestions made by Bernardo (2010) to per-
form GS for a base SI (Williams 1962), two methods were
considered for the genomic predictions of the different SI
in the present study. The first method consisted in the cal-
culation of each SI using the BLUEs across environments
as phenotypic data for the GY and PC, followed by STGS
considering the calculated SI as it was a single trait. Hereaf-
ter, this method will be called the direct method. The second
group of methods comprised firstly the genomic predictions
for the GY and the PC using separately two ST models or
jointly by means of 2TGS, followed by the calculation of
the SI considering these predictions as the phenotypic val-
ues. From now on this group of methods will be referred as
reversed methods. From the unified mixed-model and the SI
equations it becomes recognizable that the reversed meth-
ods handle information from the variance–covariance matri-
ces redundantly, similarly as it was done by Bauer and Léon
(2008) using classic BLUP approaches.
Cross‑validation for genomic predictions
In this study, four different balanced training set sizes were
considered for the cross-validation procedure; comprising
40, 80, 120 or 160 genotypes, whereas the validation set was
composed by 41 or 59 genotypes for POP-A and POP-B,
respectively. Since phenotypic data are normally generated
in an unbalanced way for the different traits within plant
breeding programs, an unbalanced data training set scenario
was additionally considered for the MT approaches. For
this scenario the training set size for one of the traits was
maintained fixed at its maximum level (i.e., 160 genotypes)
and considered the four different training set sizes for the
other(s) trait(s) included in the model. Using ST and MTGS
methods, the BLUEs across environments for GY, SEW and
PC were considered as phenotypic values for the training
set, allowing the computation of µ  and α, which were sub-
sequently used to calculate the GEBV for each trait of the
selection candidates included in the validation set. Genomic
prediction accuracies for GY and PC were calculated as the
correlation between GEBVs and the BLUEs across environ-
ments of these traits in the validation set. Prediction accura-
cies were standardized dividing them by the square root of
the heritability for each trait.
For the direct and reversed SI predictions, each train-
ing set was firstly used to estimate the G0 and P matrices.
In the case of the direct method, these estimated matrices
277
were employed to calculate the O-SI and R-SI for the train-
ing set using the BLUEs across environments as phenotypic
data for GY and PC and subsequently, by means of STGS
the GEBVs were computed for the validation set. For the
reversed method the estimated G0 and P matrices coming
from each balanced and unbalanced training set were used
to compute the different SI of the genotypes included in the
validation set. For this purpose GY and PC genomic pre-
dictions from ST- or 2TGS were used as phenotypic data
of the validation set. It was assumed that the true G0 and
P matrices would be obtained when all the genotypes of
POP-A and POP-B, respectively, were used for their esti-
mation. Then, prediction accuracies were expressed as the
correlation between the SI predictions and the SI calculated
using the true G0 and P matrices, and GY and PC BLUEs
across environments as phenotypic data for the valida-
tion set. Each prediction accuracy value presented in this
study corresponds to the average of the cross-validation
procedure repeated 1000 times. All computations were per-
formed within R environment (R Core Team 2014).
Results
Phenotypic variation, heritabilities and relationships
between traits
Phenotypic analyses for Pop-A and Pop-B have been par-
tially reported elsewhere (Miedaner et al. 2012; Wang et al.
2015) and here they are briefly summarized for GY, PC,
and SEW. BLUEs across environments reflect the high phe-
notypic variation for these three traits within POP-A and
POP-B (Table S2). High h2 values were found for GY (0.81
and 0.84 for POP-A and POP-B, respectively) and SEW
(0.87 and 0.83 for POP-A and POP-B, correspondingly),
whereas medium (0.68) and high (0.83) h2 values were esti-
mated for PC in POP-A and POP-B, respectively. Signifi-
cant phenotypic correlations were found among all pairs of
traits (Table 1). Phenotypic correlations were consistent in
terms of sign and similar in magnitude across both popu-
lations. Phenotypically, GY was negatively correlated with
PC but positively correlated with SEW, while the correla-
tion between PC and SEW was negative. For POP-A, the
strongest phenotypic correlation was found between GY
and SEW, whereas the phenotypic correlation of GY with
PC was the strongest one for POP-B. In each population,
the G0 matrices were estimated through the pre-specifica-
tion of a genomic estimated relationship matrix (Figure
S1) within an MT mixed-model, allowing subsequently the
computation of the genetic correlations between traits. In
both populations, estimates of genetic correlations resem-
ble their phenotypic counterparts in terms of sign and also
follow the same ranking of importance (Table 1).
13
278 Theor Appl Genet (2016) 129:273–287

Table 1  Phenotypic (Cor_P) Trait GY PC SEW GY PC SEW

and genetic correlations
(Cor_G) among grain yield Cor_P Cor_G
(GY, t/ha), protein content
GY – −0.454** 0.354** – −0.421 0.290
(PC, %), and single ear weight
(SEW, g) in two testcross PC −0.387** – −0.246** −0.459 – −0.161
populations (POP-A and POP- SEW 0.500** −0.227** – 0.586 −0.343 –
B) comprising each of them 220
progenies Estimated for each testcross progenies of POP-A (below diagonal) and POP-B (above diagonal) using data
for 15 (GY), ten (PC) and 12 (SEW) environments
** Significantly different from zero with P < 0.01

Genomic predictions for GY and PC 40 80 120 160

0.15
GY POP−A GY POP−B

Difference of accuracy
0.10
Predictions for GY and PC using STGS were previously
reported by Wang et al. (2015) considering 201 and 219 0.05

testcrosses from POP-A and POP-B, respectively. Briefly, 0.00

prediction accuracies for both traits were higher in POP-B −0.05

0.15
compared to POP-A, and also higher for GY than for PC PC PC 2Tb−ST
2Tufull−ST160
0.10
(Table S3). Nevertheless, in POP-B the difference between 2Tu−ST

GY and PC prediction accuracies was marginal. In the pre- 0.05

sent study, GS was performed jointly for GY and PC using 0.00

a 2T model (Table S3) and the prediction accuracies were −0.05

compared with the ones obtained by its ST counterparts
(Fig. 1). For balanced training set sizes the benefits of using
2TGS compared to STGS were marginal, if not null, for
both traits in POP-A and POP-B. Marginal improvements
were observed for PC prediction accuracies in POP-A
when the model also included GY information, whereas
these benefits were absent for POP-B. Improvements in
PC prediction accuracies were much more noticeable when
unbalanced training set sizes were used for 2TGS. In this
sense, if PC prediction is supported by maximizing GY
information to the full training set size, prediction accura-
cies are increased compared to the ones obtained by STGS
for PC. Similarly to the balanced training sets scenarios,
these improvements in PC prediction accuracies were much
more evident in POP-A than for POP-B. Nonetheless, the
benefits over STGS of 2TGS supported by maximized GY
information for PC prediction are dissipated when training
set sizes for PC are increased. Slight prediction accuracy
improvements were observed when GY predictions were
supported by maximizing PC information. Similar to the
case of PC, these improvements were much more notable
in POP-A than in POP-B and they were also dissipated as
the training set size for GY was increased. However, when
the training set size was already maximized for a particu-
lar trait, adding more information for the second trait in the
model does not appear to be much beneficial (in terms of
prediction accuracies) for the trait whose training set size is
already maximized.
After performing 2TGS, a 3TGS model adding SEW
information was also considered for POP-A (Table S4) and
prediction accuracies of this approach were compared with
40 80 120 160
Training set size
Fig. 1  Comparison between single-trait (ST) and two-trait (2T) mod-
els of the cross-validated standardized accuracy of prediction for
grain yield (GY, t/ha) and protein content (PC, %) using balanced
(2Tb) and unbalanced (2Tu) training set scenarios for population A
(POP-A) and B (POP-B). For the balanced scenario all traits share the
training set size, whereas in the unbalanced scenario only one trait
(either GY or PC) has information for 160 genotypes in the model
(namely 2Tufull) and for the remaining trait there is only information
according to the training set size. All comparisons between 2T and
ST approaches were made according to the training set size and for
the unbalanced training set scenarios the prediction accuracies for
the trait with maximized information (2Tufull) were compared with
the ones obtained by their ST counterparts using a training set size of
160 genotypes (ST160). Each point value corresponds to the average
of 1000 cross-validations
their 2T counterparts (Table 2). In general for balanced and
unbalanced training set sizes, no improvements for GY and
PC predictions accuracies were found using 3TGS over
2TGS, with the exception of GY and PC predictions sup-
ported by maximized SEW information. In this particular
case, marginal prediction accuracies improvements were
observed for GY and PC, but these also were dissipated
with increased training set sizes in the same way as the pre-
diction accuracy improvements achieved by 2TGS.
Selection methods for PC and GY
A graphical representation of the different selection strate-
gies considered in this study is included in Fig. 2. For all

13
Theor Appl Genet (2016) 129:273–287 279

Table 2  Comparison of TS VS ∆ (3Tb − 2Tb) ∆ ∆ ∆

the standardized prediction (3TufullGY − 2TufullGY) (3TufullPC − 2TufullPC) (3TufullSEW − 2Tb)
accuracies for grain yield (GY,
t/ha) and protein content (PC, GY PC GY PC GY PC GY PC
%) between the two- (2T) and
three-trait (3T) models using 40 41 −0.008 0 −0.004 0.013 −0.003 −0.007 0.041 0.047
balanced and unbalanced 80 41 −0.009 −0.002 −0.003 0.005 −0.007 −0.005 0.014 0.020
training set scenarios in
120 41 −0.006 0 −0.003 0.003 −0.005 −0.002 0.002 0.007
testcross population A
160 41 −0.003 0.001 −0.002 0.001 −0.002 0.001 −0.003 0.001

The 2T model considered GY and PC and for the 3T models the single ear weight (SEW, g) was addition-
ally taken on count. TS and VS denote training set and validation set sizes for each trait, respectively. For
the balanced (b) scenario all traits share the training set size, whereas for the unbalanced (u) scenarios only
one trait (either GY, PC or SEW) has information for 160 genotypes (namely ufull-GY, ufull-PC and ufull-SEW,
respectively) and for the remaining trait(s) there is only information according to the training set size. Each
value corresponds to the average of 1000 cross-validations

GY O-SI_(1:1) R-SI_GY
PC O-SI_(10:3) R-SI_PC
POP-A POP-B
Q1 Q2 Q1 Q2
11.0
Protein Content

10.5

10.0

9.5

Q3 Q4 Q3 Q4
9.0
6.8 7.0 7.2 7.4 7.6 7.8 6.4 6.6 6.8 7.0 7.2 7.4
Grain Yield

Fig. 2  Grain yield (GY, t/ha) and protein content (PC, %) dispersion
graphs for populations A (POP-A) and B (POP-B). Each line repre-
sents a different selection strategy but always selecting the best 10 %
of the genotypes. Vertical and horizontal lines are the culling levels
for GY and PC, respectively, and split each diagram into four differ-
ent quadrants (denoted as Q1–Q4). O-SI_(1:1) and O-SI_(10:3) refer
to the optimum selection indices (SIs) for the different GY:PC price
ratios (10:3 and 1:1), correspondingly. R-SI_GY and R-SI_PC are
the restricted SI when GY is improved while PC is maintained at the
average level or when PC is improved whereas GY is maintained at
the average level, respectively

cases, culling levels were applied to select the uppermost
10 % of the genotypes according to GY, PC, and each SI.
Culling levels for GY and PC split the dispersion graphs of
GY and PC into four quadrants (denoted as Q1, Q2, Q3 and
Q4 in Fig. 2). Quadrant Q3 corresponds to all genotypes
which are culled because of their low GY and PC levels.
When independent culling levels for GY and PC are used
and both traits are intended to be simultaneously improved,
selection for one trait would completely impede the use
of the same selection criteria for the second trait, which
is reflected by an empty Q2 quadrant. Therefore, no geno-
types could be selected for high GY and PC simultaneously
using independent culling levels. Furthermore, if the GY
and PC culling levels are considered as part of a tandem
selection strategy, the improvement of one of the traits dur-
ing the first generation(s) would limit the improvement of
the second trait during the next generation(s) (considered
as a breeding cost for the second trait, when its economic
importance is ignored). Quadrants Q4 and Q1 represent
the breeding costs in GY and PC, respectively, when these
traits are selected during the late generations of a tandem
selection program which aims to improve GY and PC
simultaneously. However, an SI such as the O-SI can partly
reduce these breeding costs and also allow the inclusion of
economic information. If both traits have the same scale,
the most balanced case of simultaneous selection for high
GY and PC values would be expected for the O-SI, when
both traits have the same economic importance. The O-SI
would allow the selection of genotypes that maximize GY
and PC simultaneously or also of genotypes which com-
pensate the not so high values for one trait with high values
for the second trait. When the GY acquires 10:3 times the
economic importance of PC, it can be seen that the curve
of the O-SI approaches towards the culling level for GY,
because genotypes maximizing GY would be preferred for
selection. The curves representing the R-SIs (R-SI_GY and

13
280
1.0
GY POP-A
PC
0.8
O-SI_(1:1)
O-SI_(10:3)
R-SI_GY
Accuracy
0.6 R-SI_PC
0.4
0.2
0.0
40 80 120
Training set size
Fig. 3  Relationship between training set size and the cross-validated
accuracies of prediction for single-trait genomic selection of grain
yield (GY, t/ha), protein content (PC, %) and the different selection
indices (SIs) in populations A (POP-A) and B (POP-B). O-SI_(1:1)
and O-SI_(10:3) refer to the optimum selection indices for the dif-
R-SI_PC) are located towards the culling level of the trait
which is intended to be improved.
Genomic predictions for SIs
The accuracies of genomic predictions for the different
SIs of GY and PC according to the direct method are rep-
resented in Fig. 3. Similar to the STGS for GY and PC,
when the training set size was increased, prediction accu-
racies improved for all SIs and independent of the train-
ing set size, prediction accuracies for all SIs were higher
in POP-B than in POP-A. In both populations, the highest
prediction accuracies were obtained for the O-SI with a
10:3 GY:PC price ratio followed by the ones observed for
the R-SI when GY improvement was the breeding goal.
In POP-A, the O-SI with a 1:1 GY:PC price ratio and the
R-SI for PC improvement had the lowest prediction accura-
cies. For POP-B the O-SI with a 1:1 GY:PC price ratio had
also the lowest SIs prediction accuracies, but the R-SI for
PC improvement had prediction accuracies similar to the
ones obtained for an R-SI when GY improvement was the
breeding goal. All these differences in prediction accura-
cies between the different SIs were more evident in POP-A
than in POP-B.
A group of reversed prediction methods reconstruct-
ing each SI from ST- and 2TGS predictions for GY and PC
was implemented and then compared to the direct method of
SI prediction to see if improvements in prediction accuracy
could be achieved by the reversed methods (Table 3). Using
reversed methods based on ST- and 2TGS with balanced
13
Theor Appl Genet (2016) 129:273–287
POP-B
160 40 80 120 160
ferent GY:PC price ratios (1:1 and 10:3), respectively. R-SI_GY and
R-SI_PC are the restricted selection index when GY is improved
while PC is maintained at the average level or when PC is improved,
whereas GY is maintained at the average level, correspondingly. Each
point value corresponds to the average of 1000 cross-validations
training sets for GY and PC, the improvements were mar-
ginal for some of the SIs and tended to disappear when the
training set size increased. In POP-A, marginal improvements
were observed for the O-SI with a 1:1 GY:PC price ratio
reconstructed from ST predictions for GY and PC, and also
for the R-SI for GY improvement calculated from both ST-
and 2TGS predictions. For POP-B, marginal improvements
could be only observed for the R-SI for PC improvement
reconstructed using 2TGS predictions. Nevertheless, in some
cases the prediction accuracies of the reversed methods were
slightly lower than the ones obtained by their direct counter-
parts, being the most biased case the predictions of the R-SI
for PC improvement based on STGS predictions for GY and
PC in POP-A. However, when the 2TGS predictions for GY
and PC using unbalanced training set sizes were employed
in the reversed method, much more prediction accuracy
improvements could be observed. When 2TGS predictions
were obtained by maximizing GY information to the full
training set size, notable improvements were observed par-
ticularly for the O-SI with a 1:1 GY:PC price ratio in POP-A,
whereas in both populations substantial improvements were
achieved for the O-SI with a 10:3 price ratio and the R-SI
for GY improvement. If 2TGS predictions were acquired by
means of maximizing PC information to the full training set
size, considerable improvements were found for the R-SI for
PC improvement in both populations followed by improve-
ments for the O-SIs with 1:1 and 10:3 GY:PC price ratios in
POP-B and A, respectively. Nonetheless, similar to the case
of balanced training set sizes, the advantages of the reversed
prediction methods using unbalanced training set sizes tended
Theor Appl Genet (2016) 129:273–287 281

Table 3  Comparison of the cross-validated accuracy of genomic predictions for the different selection indices (SI) between a direct prediction
method (DST) and four different reversed prediction methods (R_) in two rye testcross populations, POP-A and POP-B
Population TS VS ∆ (R_ST-DST) ∆ (R_MTb-DST)
O-SI_(1:1) O-SI_(10:3) R-SI_GY R-SI_PC O-SI_(1:1) O-SI_(10:3) R-SI_GY R-SI_PC

POP-A 40 41 0.013 0.006 0.017 −0.013 0.005 0.009 0.018 0.007

80 41 0.009 0.007 0.014 −0.015 0.005 0.010 0.016 0.009
120 41 0.008 0.003 0.008 −0.015 0.006 0.008 0.013 0.008
160 41 0.004 −0.001 0.003 −0.013 0.004 0.006 0.010 0.008
POP-B 40 59 0.004 0.004 0.005 −0.001 0.001 0.000 0.003 0.015
80 59 0.000 0.000 −0.001 −0.002 −0.002 −0.003 0.003 0.005
120 59 −0.002 −0.002 −0.003 −0.003 −0.001 −0.002 0.001 0.001
160 59 −0.002 −0.001 −0.002 −0.003 −0.001 −0.002 0.000 0.000
Population TS VS ∆ (R_MTufullGY-DST) ∆ (R_MTufullPC-DST)
O-SI_(1:1) O-SI_(10:3) R-SI_GY R-SI_PC O-SI_(1:1) O-SI_(10:3) R-SI_GY R-SI_PC

POP-A 40 41 0.055 0.145 0.141 0.010 −0.005 0.031 −0.004 0.099

80 41 0.027 0.071 0.073 0.011 0.002 0.017 0.009 0.055
120 41 0.013 0.029 0.033 0.009 0.007 0.009 0.011 0.031
160 41 0.004 0.006 0.010 0.008 0.004 0.006 0.010 0.008
POP-B 40 59 0.005 0.118 0.133 0.004 0.090 −0.004 −0.040 0.134
80 59 0.006 0.044 0.049 0.005 0.033 −0.005 −0.002 0.051
120 59 0.002 0.014 0.016 0.002 0.012 −0.002 0.001 0.018
160 59 −0.001 −0.002 0.000 0.000 −0.001 −0.002 0.000 0.000

GY and PC denote the traits grain yield (t/ha) and protein content (%), correspondingly. O-SI_(1:1) and O-SI_(10:3) refer to the optimum SIs
for the different GY:PC price ratios (10:3 and 1:1), respectively. R-SI_GY and R-SI_PC are the restricted SIs when GY or PC, correspondingly,
are intended to be improved and the remaining trait is maintained at the average level. DST is the implementation of genomic selection using
directly each SI as a single-trait (ST) to train the model. R_ST corresponds to the reversed method using GY and PC predictions coming from
two independent ST methods. R_MTb and R_MTu are the reversed methods using GY and PC predictions originated by two-trait (2T) models
in balanced and unbalanced scenarios, respectively. TS and VS denote training set and validation set sizes, respectively. For the balanced sce-
nario all traits share the training set size, whereas for the unbalanced scenario only one trait (either GY or PC) has information for 160 genotypes
in the model (namely MTufull-GY or MTufull-PC, respectively) and for the remaining trait there is only information according to the training set
size. Comparisons correspond to the difference in prediction accuracies between R_ and DST methods. Each value corresponds to the average of
1000 cross-validations

to banish when the training set size increased, with the excep-
tion of the R-SI for GY improvement when PC information
was maximized to the full training set size. For this specific
case, reversed predictions for the R-SI for GY improve-
ment appeared to be inferior compared to the direct method
in small training set sizes, but as the training set size for GY
was increased, the reversed method tended to reach, and par-
ticularly in POP-A slightly overcome, the prediction accuracy
level of the direct method.
Discussion
MTGS can be performed in order to achieve more
cycles of selection per unit of time
In plant science and compared to STGS studies, the appli-
cation of MTGS methods aiming to benefit from the
information contained in correlated indicator traits is rela-
tively scarce, with only three studies to the date published
using real/experimental plant breeding data (Bao et al.
2015; Jia and Jannink 2012; Rutkoski et al. 2012). Jia and
Jannink (2012) benefited from phenotype imputation using
MTGS applied to pine breeding data. In this sense, com-
pared to STGS, it was possible for them to obtain a 60 %
increase in prediction accuracy for a partially masked trait
when all the genotypes have been measured for another
particular trait. Similarly, Rutkoski et al. (2012) found sub-
stantial prediction accuracy improvements and also higher
gains per cycle over STGS when the kernel quality index
was additionally included as predictor variable within
an MTGS model for deoxynivalenol toxin levels predic-
tion in wheat. Nonetheless, the results from both studies
imply that, in practice, the correlated indicator trait should
be first measured for the selection candidates before per-
forming MTGS for a specific target trait. In consequence,

13
282
a cost-effective pre-requisite for the implementation of
these MTGS methods is that the evaluation of the corre-
lated indicator trait must be less costly than phenotyping
the target trait directly. In addition, such MTGS implemen-
tations do not allow to accelerate the breeding process by
reaching more cycles of selection per unit of time (Rut-
koski et al. 2012), which is one of the main advantages of
STGS (Longin et al. 2015; Rutkoski et al. 2012). Recently,
Bao et al. (2015) used different MTGS models concerning
various combinations of traits related to sudden death syn-
drome (SDS) resistance in soybean; however, they found
no significant benefits when shifting from STGS to MTGS
models. We presume that these past findings have some-
how discouraged the plant breeding community to imple-
ment MTGS as a common routine within plant breeding
programs. Nevertheless, this situation also stimulated us to
perform MT genomic predictions for the selection candi-
dates only relying on trait information contained within the
training sets. Therefore, the positive results achieved in our
study, and discussed in the following manuscript sections,
indicate that MTGS could also be implemented in order to
achieve more cycles of selection by unit of time in the same
way as STGS. The theoretical and practical implications of
our positive findings for the plant breeding community in
general and for the improvement of GY and PC in rye for
feeding purposes are discussed.
Correlations among GY, PC and SEW
Jia and Jannink (2012) concluded that genetic correla-
tions between traits are the basis for the benefits of MTGS.
Therefore, in order to avoid the application of MTGS
using an unfavorable scenario for their benefits, we com-
puted first the phenotypic and genetic correlations between
GY, PC and SEW. Phenotypic correlations among these
traits were previously reported by Miedaner et al. (2012)
for POP-A and POP-B using a smaller number of environ-
ments. The phenotypic correlations found in the present
study (Table 1) are coincident in sign and similar in mag-
nitude with the previous reports. Negative phenotypic and
genetic correlations between GY and PC were found in
POP-A and POP-B. Since starch content (SC) is positively
correlated with GY (r = 0.208 and 0.356 at the phenotypic
level for POP-A and POP-B, respectively) and negatively
correlated to PC (phenotypic correlations of −0.601 and
−0.809 for POP-A and POP-B, correspondingly), the nega-
tive relationship between GY and PC could be partly medi-
ated by the SC. In this sense, a higher SC would increase
GY but would simultaneously decrease PC in a competi-
tive way, since both correspond to seed storage constitu-
ents (Kindred et al. 2008; Perez et al. 1996). Positive phe-
notypic and genetic correlations were found between GY
and SEW in both populations, which is expected because
13
Theor Appl Genet (2016) 129:273–287
SEW corresponds to one important GY component in cere-
als (Lin and Allaire 1977; Pinto et al. 2002; Yoshira et al.
2000). For both populations, negative phenotypic and
genetic correlations between PC and SEW were observed.
However, mediating effects of other traits for this relation-
ship were not so clear and the possible causes of the nega-
tive correlation between SEW and PC are beyond the scope
of the present study.
Trait heritabilities, data structure orthogonality
and genetic correlations are the main driving forces
influencing the benefits of MTGS over STGS for GY
and PC within the two rye testcross populations
In practice, if a target trait is difficult to measure with pre-
cision, the errors of its measurement may so reduce herit-
ability that indirect selection using a genetically correlated
indicator trait with higher heritability becomes advanta-
geous (Falconer and MacKay 1996). Consequently, it has
been shown in past studies using simulated data that MTGS
is beneficial compared to STGS mostly when the genomic
predictions of a trait with relatively low heritability are sup-
ported by using information from a genetically correlated
trait with higher heritability (Guo et al. 2014; Hayashi and
Iwata 2013; Jia and Jannink 2012). In parallel, neither the
changes in heritability of the correlated trait supporting the
predictions nor changes in the genetic correlations between
traits influence the prediction accuracies for a highly herita-
ble trait (Jia and Jannink 2012). Therefore, not much bene-
fit should be expected for highly heritable traits when shift-
ing from STGS to MTGS. The results of the present study
using experimental rye data agree with these past findings.
High heritability values for GY and PC were estimated in
POP-B (h2 = 0.84 and 0.83 for GY and PC, respectively);
hence the benefits from 2TGS over STGS were marginal,
if not null, for both traits within this particular population
(Fig. 1). However, 2TGS was more beneficial over STGS
for PC prediction in POP-A, where the heritability of PC
(h2 = 0.68) was lower than the one of GY (h2 = 0.81), and
these improvements were even more pronounced within
unbalanced training sets, when the GY was set to the maxi-
mum training set size. In addition, using unbalanced train-
ing set sizes, some prediction accuracy improvements for
GY were observed by shifting from STGS to 2TGS when
the information for this trait was partially available and the
PC information was maximized to the full training set size
in POP-A (Fig. 1). Guo et al. (2014) observed using simu-
lated data that when the phenotypic information is partially
and completely available for the target and the genetically
correlated indicator trait, respectively, a case analogous to
the unbalanced training sets in the present study, MTGS
performs much better than STGS, especially for a target
trait with lower heritability. In consequence, the superiority
Theor Appl Genet (2016) 129:273–287
of 2TGS over STGS using unbalanced training set sizes in
the present study was dissipated as the training set size for
the partially available target trait increased (Fig. 1).
Genetic correlations also influence the relative superior-
ity of MTGS over STGS. From studies on simulated data,
when one trait with high heritability is supporting MTGS
predictions for a trait with lower heritability, the more
genetically correlated are these two traits, the higher the
prediction accuracies would be (Calus and Veerkamp 2011;
Hayashi and Iwata 2013; Jia and Jannink 2012). Further-
more, Bao et al. (2015) found no significant prediction
accuracy improvements over STGS using MTGS and they
attributed this lack of substantial improvement to the weak
correlations among traits. In the present study, the benefits
of 2T- over STGS were more noticeable for POP-A than
for POP-B (Fig. 1). Even though the absolute value of the
phenotypic correlation between GY and PC was higher for
POP-B than for POP-A, the genetic correlation between
GY and PC was slightly more pronounced for POP-A than
for POP-B (Table 1). In this sense, the genetic correlations
between GY and PC could also partly explain why the ben-
efits of 2T- over STGS were more evident for POP-A than
POP-B.
In our study, POP-B had a higher marker density com-
pared to POP-A, with 584 and 394 DArT markers for each
population, respectively. In general, simulation studies on
STGS have shown that an increased marker density has a
positive effect on the prediction accuracy of GS (Meuwis-
sen 2009; Solberg et al. 2008). However, we did not directly
evaluate the prediction accuracies reached by MTGS itself,
but the difference in prediction accuracies compared to
STGS models (Fig. 1; Table 2); hence, we anticipate that
the influence of marker density mentioned in past simula-
tion studies is not relevant on our results. Nevertheless, to
our knowledge the influence of marker density on the supe-
riority of MTGS over STGS has not been deeply studied
so far. To rule out this potential confounding effect caused
by the different marker densities of POP-A and POP-B, we
performed ST- and 2TGS for GY and PC in POP-B using
a balanced training set scenario but now jointly sampling
markers (with a sample size of 394 DArT markers) and gen-
otypes (according to each training set size) for each cross-
validation run. Differences in prediction accuracy of MTGS
over STGS obtained by this approach were extremely close
to the original ones (data not shown). Therefore, we expect
that the main driving forces explaining the different perfor-
mances of MTGS over STGS between POP-A and POP-B
(Fig. 1) are due to the differences in heritabilities (Table
S2) and genetic correlations (Table 1) of both populations.
Nonetheless, studies on MTGS using different marker den-
sities should be performed in the near future, additionally
considering factors such as the genetic architecture and the
level of phenotyping unbalance within training sets for the
283
predicted traits. These future studies would allow a better
understanding of the marker density effect on the superior-
ity of MTGS over STGS and also to potentially generalize
our findings to other scenarios.
In summary, our recent findings using experimental rye
breeding data plus the results from past simulation (Calus
and Veerkamp 2011; Guo et al. 2014; Hayashi and Iwata
2013; Jia and Jannink 2012) and experimental plant data
(Bao et al. 2015) studies allow a practical understand-
ing about the ideal conditions needed to exploit the ben-
efits of MTGS over STGS. In this sense, an ideal scenario
for plant breeders to take advantage of MTGS would be
when the predictions of a low-heritability trait with scarce
phenotypic records are supported by a genetically corre-
lated highly heritable indicator trait which was massively
measured.
Practical implications of MTGS for PC and GY in rye
According to our results using balanced training set sce-
narios, only PC in POP-A benefited from 2T-over STGS
(Fig. 1). Although these improvements in prediction accu-
racy were not pronounced, they imply that when PC has
a relatively low heritability, its prediction will be slightly
more accurate than performing STGS if a 2TGS model
is applied using GY as a highly heritable genetically cor-
related indicator trait. In addition, GY and PC of POP-A
benefited notably from MTGS when the phenotypic infor-
mation within the training set for the target and indicator
trait was partially and completely available, respectively
(Fig. 1). In general, due to the high economic importance
of GY in cereal crops, plant breeding programs would tend
to handle more phenotypic records for GY than for any
other trait. Therefore, taking advantage of the benefits of
2TGS over STGS using unbalanced training sets would be
more realistic for PC than for GY. However, since the SEW
is an important component determining GY in cereals (Lin
and Allaire 1977; Pinto et al. 2002; Yoshira et al. 2000) and
both traits were positively correlated in the present study
(Table 1), we further examined the potential benefits over
STGS of using the SEW as indicator trait to predict GY by
means of 2TGS (Table S5). Pronounced benefits in terms of
GY prediction accuracies of 2TGS over STGS were only
observed in the unbalanced training set scenario with maxi-
mized SEW information for POP-A. We presume that the
lack of benefits of 2TGS over STGS within balanced train-
ing set scenarios was due to the high and similar heritabil-
ity estimates for GY and SEW in both rye testcross pop-
ulations (Table S2). Additionally, the almost null benefits
from 2TGS over STGS using unbalanced training sets in
POP-B could be mainly a consequence of the substantially
lower genetic correlation between GY and SEW compared
to the one of POP-A (Table 1) and secondly, because SEW
13
284
was slightly less-heritable than GY in POP-B (Table S2).
Nevertheless, even though phenotyping for SEW could be
theoretically less time- and resource-consuming than meas-
uring GY in field trials, the plausibility of this last MTGS
strategy for GY prediction using the SEW as correlated
indicator trait should be evaluated according to the optimal
allocation of resources between GY and SEW phenotyping.
We consider that, in general, the search for realistic and
cost-effective strategies for the implementation of MTGS
in plant breeding programs should be ultimately conducted,
similarly as it has been already done for STGS implemen-
tation for GY prediction in cereal breeding (Endelman
et al. 2013; Longin et al. 2015).
Optimizing the number of traits handled by MTGS
models
Models for MTGS have been rarely fitted in the past using
more than two traits simultaneously. Tsuruta et al. (2011)
performed MTGS using a model which simultaneously con-
sidered 18 traits of US Holstein cattle. They argued that the
prediction accuracies are increased when switching from
STGS to MTGS but that the increase depends on the trait
being predicted. In contrast, Bao et al. (2015) did not find
significant benefits when shifting from STGS to an MTGS
model which simultaneously considered four traits related
to SDS resistance in soybean. However, these studies did
not discuss the benefits and limitations of progressively
adding additional traits to an MTGS model which already
performed well with a few traits. Furthermore, MT meth-
ods are known for being more time and computer resource
consuming than their ST counterparts, and this limitation
is expected to become more serious when the number of
phenotypic records or traits increases (Bauer and Léon
2008; Calus and Veerkamp 2011; Hayashi and Iwata 2013).
Probably, this is another of the reasons why MTGS has not
been routinely adopted until now. In consequence, an opti-
mal equilibrium between prediction accuracy improvement
and computation efficiency should be somehow investi-
gated when additional traits are considered within the pre-
dictive model. In the present study, GY and PC prediction
accuracy improvements were only observed for an MTGS
model including GY, PC and SEW using a training set with
maximized SEW information in POP-A (Table 2). There-
fore, one of the advantages of 3T-over 2TGS is that it can
simultaneously improve the prediction accuracies for two
traits with scarce phenotypic records when the information
for a third correlated trait is completely available. Never-
theless, some cautions should be taken on count when
shifting from 2T-to 3TGS models. Besides enhancing the
time and computer resources requirements, when more cor-
related traits are added to the MTGS model, the probability
of having colinearity problems could increase, which would
13
Theor Appl Genet (2016) 129:273–287
impose difficulties such as convergence problems into the
solving of the mixed-model equations. Computational bur-
den would be also enhanced when unbalanced datasets are
handled by the mixed model equations. Moreover, it should
be also expected that the contributions of the new traits
which are added to the model would tend asymptotically
towards zero; hence the more traits an MTGS model has
the lesser would be the contribution of a new trait added
to the model. In addition, the interpretation of models with
more traits becomes more complex. Last but not least, in
the present study the prediction accuracies obtained for a
trait with scarce phenotypic records available using 2TGS
and an indicator trait with complete information at hand
(Tables S3 and S5) are in general higher, if not similar, than
the ones obtained by 3TGS with partial information for the
same two traits and complete information available for a
third correlated trait (Table S4). According to these results
for GY and PC in rye, if the aim is to improve the predic-
tion accuracy of only one particular trait, 2TGS should be
sufficient for that purpose.
Predicting SIs for GY and PC in rye
There are several possible selection procedures to maxi-
mize the economic value of animals or plants, including
tandem selection, independent culling levels or simulta-
neous selection by means of an index. The SI method is
expected to give the most rapid improvement of the eco-
nomic value (Falconer and MacKay 1996), and therefore
the application of GS to predict different SIs could poten-
tially break the gap between GS and MT improvement. In
the present study, firstly, a direct method for SI prediction
was presented, in which each SI was treated as a single
trait within the GS model. Four SIs were predicted using
this method and different prediction accuracy levels were
achieved for each of them (Fig. 3). Since the b values deter-
mine the contribution of each trait to the SI, we expected
that the relative importance of each trait contributing to a
particular SI (see bGY:bPC ratio in Table S6) and also GY
and PC prediction accuracies influenced the prediction
accuracy of the SI. However, this trend was not clearly
observed. Nevertheless, although the differences in herita-
bilities were not so high among the different SI (Table S6),
in general, the higher the heritability of the SI, the better
was the prediction accuracy achieved (Fig. 3). The herit-
ability of a sum of traits not only include the phenotypic
and genetic variances of the traits, but also the co-variances
among them; hence, it would not necessarily match the
average of the heritability of the traits composing the index
(Openshaw and Hadley 1984). This makes the heritability
of an SI difficult to interpret, but also suggests that the co-
variances between traits play some role influencing the pre-
diction accuracies of the SI.
Theor Appl Genet (2016) 129:273–287
In addition, a group of reversed prediction methods,
constructing each SI based on ST- and 2TGS predictions
for GY and PC, were computed and then compared to
the direct method (Table 3). Bauer and Léon (2008) con-
structed an O-SI for two simulated traits with a 1:1 price
ratio by means of classic ST- and MT-BLUP approaches
in a similar way as the reversed methods using balanced
training sets of the present study. In general, constructing
the SI based on MT-BLUPs allowed them to obtain better
selection responses than using ST-BLUPs, especially when
two traits were negatively correlated. However, in terms of
prediction accuracies and compared to the direct SI predic-
tion method, substantial improvements were only observed
for some SIs predicted by means of the reversed methods
based on 2TGS predictions using unbalanced training sets
in our study (Table 3). One of the disadvantages of the
direct method is that it can only handle balanced pheno-
typic information and in consequence cannot exploit the
benefits of MTGS using unbalanced training sets for dif-
ferent traits. Nonetheless, these improvements in prediction
accuracy were not observed for all the SIs and we presume
that this inconsistency is partially related to the b values,
although it is beyond our faculties to demonstrate it. Never-
theless, as a general recommendation, when balanced phe-
notypic information is available, the direct method should
be preferred for SI genomic prediction, being simpler and
more efficient than the reversed methods.
Conclusion
This is the first study using experimental plant breeding data
that (1) intensively discusses the practical implications of
exploiting the information of correlated traits by means of
MTGS and (2) also applies GS tools for different SI to con-
front an undesired correlation between two traits which are
intended to be bred in rye for feeding purposes. From the
results, it can be concluded that most of the findings from past
studies on MTGS using simulated data also hold truth for GY
and PC data in rye. Then, the ideal scenario to exploit the ben-
efits of MTGS would be when the predictions of a relatively
low heritable trait with scarce phenotypic records are sup-
ported by a genetically correlated highly heritable indicator
trait which was measured on a large scale. In a plant breed-
ing context, this ideal scenario is more realistic for PC than
for GY. In the present study, GS was performed only relying
on trait information contained within the training sets, which
indicates using experimental plant breeding data that MTGS
could also be implemented in order to achieve more cycles
of selection by unit of time in the same way as STGS. Even
though, it is not expected that the difference in marker den-
sity between the two testcross rye populations influenced
the superiority of MTGS over STGS, these results should be
285
validated by means of simulations in order to be extrapolated
to other situations. In addition, although 3TGS could allow the
simultaneous improvement of the prediction accuracies for
two traits with partial information, if the aim is to exclusively
improve the prediction accuracy for a particular single trait, a
2TGS model is expected to be more suited for that purpose. In
general, for balanced phenotypic information, we recommend
to perform GS considering the SI as a single trait, this method
being a simple, direct and efficient way of prediction. Last but
not least, studies comparing the performances of the GS meth-
ods suggested for a base SI by Bernardo (2010), and presented
for the O-SI and R-SI in our study, with the ones achieved by
alternative SI genomic prediction methods (Cerón-Rojas et al.
2008; Dekkers 2007), would also be required. We expect that
the joint consideration of past simulated and real/experimen-
tal data studies plus our recent findings allow a better under-
standing of MT improvement by means of GS and opens the
door to future studies intensively discussing how to implement
these methods in a cost-effective manner by plant breeding
institutions and companies.
Author contribution statement YW and AS wrote the
manuscript. YW, AS and YZ performed calculations and
constructed figures and tables. TM and PW supervised the
acquisition of phenotypic data. YW, AS, YZ and JCR inter-
preted the results. YZ, JCR, TM and PW helped to improve
the manuscript.
Acknowledgments This research was conducted within the pro-
ject “Erweiterung der genetischen Basis von Hybridroggen für
Korn- und Biomasseleistung sowie Trockenheitstoleranz mittels
Mehrlinienkartierung und DH-Technik” financially supported by the
German Federal Ministry of Food and Agriculture via the “Fachagen-
tur Nachwachsende Rohstoffe e.V.”, Gülzow, Germany (Grant ID:
22021711).
Compliance with ethical standards
Conflict of interest The authors declare that they have no conflict
of interest.
Ethical statement The experiments were performed according to
the current laws of Germany.
References
Albrecht T, Wimmer V, Auinger HJ, Erbe M, Knaak C, Ouzunova M,
Simianer H, Schön CC (2011) Genome-based prediction of test-
cross values in maize. Theor Appl Genet 123:339–350
Bao Y, Kurle JE, Anderson G, Yong ND (2015) Association map-
ping and genomic prediction for resistance to sudden death
syndrome in early maturing soybean germplasm. Mol Breed
35:128
Bauer A, Léon J (2008) Multiple-trait breeding values for paren-
tal selection in self-pollinating crops. Theor Appl Genet
116:235–242
13
286
Bernardo RN (2010) Breeding for quantitative traits in plants.
Stemma Press, Woodbury
Butler DG, Cullis BR, Gilmour AR, Gogel B (2009) ASReml-R refer-
ence manual. The State of Queensland, Department of Primary
Industries and Fisheries, Brisbane
Calus MPL, Veerkamp RF (2011) Accuracy of multi-trait genomic
selection using different methods. Genet Sel Evol 43:26
Cerón-Rojas JJ, Sahagún-Castellanos J, Castillo-González F, Santacruz-
Varela A, Crossa J (2008) A restricted selection index method
based on eigenanalysis. J Agric Biol Environ Stat 13:440–457
Christensen OF, Madsen P, Nielsen B, Ostersen T, Su G (2012)
Single-step methods for genomic evaluation in pigs. Animal
6:1565–1571
Crossa J, de Los Campos G, Pérez P, Gianola D, Burgueño J, Araus
JL, Makumbi D, Singh RP, Dreisigacker S, Yan J, Arief V, Ban-
ziger M, Braun HJ (2010) Prediction of genetic values of quan-
titative traits in plant breeding using pedigree and molecular
markers. Genetics 186:713–724
Dekkers JCM (2007) Prediction of response to marker-assisted and
genomic selection using selection index theory. J Anim Breed
Genet 124:331–341
Dolan DJ, Stuthman DD, Kolb FL, Hewings AD (1996) Multiple trait
selection in a recurrent selection population in oat (Avena sativa
L.). Crop Sci 36:1207–1211
Eagles HA, Frey KJ (1974) Expected and actual gains in eco-
nomic value of oat lines from five selection methods. Crop Sci
14:861–864
Elgin JH, Hill RR, Zeiders KE (1970) Comparison of four meth-
ods of multiple trait selection for five traits in alfalfa. Crop Sci
10:190–193
Endelman JB, Atlin GN, Beyene Y, Semagn K, Zhang X, Sorrells ME,
Jannink J-L (2013) Optimal design of preliminary yield trials
with genome-wide markers. Crop Sci 54:48–59
Falconer DS, MacKay TFC (1996) Introduction to quantitative genet-
ics, 4th edn. Ronald Press Company, New York
Guo G, Zhao F, Wang Y, Zhang Y, Du L, Su G (2014) Comparison of
single-trait and multiple-trait genomic prediction models. BMC
Genet 15:30
Hayashi T, Iwata H (2013) A Bayesian method and its variational
approximation for prediction of genomic breeding values in mul-
tiple traits. BMC Bioinformatics 14:34
Hazel LN (1943) The genetic basis for constructing selection indexes.
Genetics 28:476–490
Hazel LN, Lush JL (1942) The efficiency of three methods of selec-
tion. J Hered 33:393–399
Heffner EL, Jannink JL, Iwata H, Souza E, Sorrells ME (2011)
Genomic selection accuracy for grain quality traits in biparental
wheat populations. Crop Sci 51:2597–2606
Henderson CR (1975) Best linear unbiased estimation and prediction
under a selection model. Biometrics 31:423–447
Henderson CR, Quass RL (1976) Multiple trait evaluation using rela-
tives’ records. J Anim Sci 43:1188–1197
Heslot N, Yang HP, Sorrells ME, Jannink JL (2012) Genomic selection
in plant breeding: a comparison of models. Crop Sci 52:146–160
Hofheinz N, Borchardt D, Weissleder K, Frisch M (2012) Genome-
based prediction of test cross performance in two subsequent
breeding cycles. Theor Appl Genet 125:1639–1645
Holbrook CC, Burton JW, Carter TE (1989) Evaluation of recurrent
restricted index selection for increasing yield while holding seed
protein constant in soybean. Crop Sci 29:324–329
Hübner M, Wilde P, Schmiedchen B, Dopierala P, Gowda M, Reif
JC, Miedaner T (2013) Hybrid rye performance under natural
drought stress in Europe. Theor Appl Genet 126:475–482
Jarquín D, Kocak K, Posadas L, Hyma K, Jedlicka J, Graef G, Lorenz
A (2014) Genotyping by sequencing for genomic prediction in a
soybean breeding population. BMC Genomics 15:740
13
Theor Appl Genet (2016) 129:273–287
Jia Y, Jannink JL (2012) Multiple-trait genomic selection meth-
ods increase genetic value prediction accuracy. Genetics
192:1513–1522
Kempthorne O, Nordskog AW (1959) Restricted selection indices.
Biometrics 15:10–19
Kindred DR, Verhoeven TMO, Weightman RM, Swanston JS, Agu
RC, Brosnan JM, Sylvester-Bradley R (2008) Effects of vari-
ety and fertiliser nitrogen on alcohol yield, grain yield, starch
and protein content, and protein composition of winter wheat. J
Cereal Sci 48:46–57
Lin C, Allaire F (1977) Heritability of a linear combination of traits.
Theor Appl Genet 51:1–3
Liu T, Qu H, Luo C, Li X, Shu D, Lund MS, Su G (2014) Genomic
selection for the improvement of antibody response to newcas-
tle disease and avian influenza virus in chickens. PLoS One.
doi:10.1371/journal.pone.0112685
Longin CFH, Mi X, Würschum T (2015) Genomic selection in wheat:
optimum allocation of test resources and comparison of breed-
ing strategies for line and hybrid breeding. Theor Appl Genet
128:1297–1306
Meuwissen THE (2009) Accuracy of breeding values of ‘unrelated’
individuals predicted by dense SNP genotyping. Genet Sel Evol
41:35
Meuwissen THE, Hayes BJ, Goddard ME (2001) Prediction of total
genetic value using genome-wide dense marker maps. Genetics
157:1819–1829
Miedaner T, Hübner M, Korzun V, Schmiedchen B, Bauer E,
Haseneyer G, Wilde P, Reif JC (2012) Genetic architecture of
complex agronomic traits examined in two testcross populations
of rye (Secale cereale L.). BMC Genomics 13:706
Openshaw SJ, Hadley HH (1984) Selection indexes to modify protein
concentration of soybean seeds. Crop Sci 24:1–4
Perez C, Juliano BO, Liboon SP, Alcantara JM, Cassman KG (1996)
Effects of late nitrogen fertilizer application on head rice
yield, protein content, and grain quality of rice. Cereal Chem
73:556–560
Piepho HP, Möhring J (2007) Computing heritability and selec-
tion response from unbalanced plant breeding trials. Genetics
177:1881–1888
Pinto RJB, Alvarez JB, Martínet LM (2002) Preliminary evaluation
of grain yield components in hexaploid tritordeum. Crop Breed
Appl Biotechnol 2:213–218
R Core Team (2014) R: A language and environment for statistical
computing. R Foundation for Statistical Computing, Vienna,
Austria. ISBN 3-900051-07-0. http://www.R-project.org
Reif JC, Hamrit S, Heckenberger M, Schipprack W, Maurer HP, Bohn
M, Melchinger AE (2005) Trends in genetic diversity among
European maize cultivars and their parental components during
the past 50 years. Theor Appl Genet 111:838–845
Rutkoski J, Benson J, Jia Y, Brown-Guedira G, Jannink JL, Sorrells
ME (2012) Evaluation of genomic prediction methods for Fusar-
ium head blight resistance in wheat. Plant Genome 5:51–61
Searle SR (1982) Matrix algebra useful for statistics. Wiley, New York
Shu Y, Yu D, Wang D, Bai X, Zhu Y, Guo C (2013) Genomic selection
of seed weight based on low-density SCAR markers in soybean.
Genet Mol Res 12:2178–2188
Smith HF (1936) A discriminant function for plant selection. Ann
Eugen 7:240–250
Solberg TR, Sonesson AK, Woolliams JA, Meuwissen T (2008)
Genomic selection using different marker types and densities. J
Anim Sci 86:2447–2454
Spindel J, Begum H, Akdemir D, Virk P, Collard B, Redoña E, Atlin
G, Jannink JL, McCouch SR (2015) Genomic selection and
association mapping in rice (Oryza sativa): effect of trait genetic
architecture, training population composition, marker number
and statistical model on accuracy of rice genomic selection in
Theor Appl Genet (2016) 129:273–287
elite, tropical rice breeding lines. PLoS Genet. doi:10.1371/jour-
nal.pgen.1004982
Suwantaradon K, Eberhart SA, Mock JJ, Owens JC, Guthrie WD
(1975) Index selection for several agronomic traits in the BSSS2
maize population. Crop Sci 15:827–833
Tsuruta S, Misztal I, Aguilar I, Lawlor TJ (2011) Multiple-trait
genomic evaluation of linear type traits using genomic and phe-
notypic data in US Holsteins. J Dairy Sci 94:4198–4204
Wang Y, Mette MF, Miedaner T, Gottwald M, Wilde P, Reif JC, Zhao
Y (2014) The accuracy of prediction of genomic selection in elite
hybrid rye populations surpasses the accuracy of marker-assisted
selection and is equally augmented by multiple field evaluation
locations and test years. BMC Genomics 15:556
Wang Y, Mette MF, Miedaner T, Wilde P, Reif JC, Zhao Y (2015) First
insights into the genotype–phenotype map of phenotypic stabil-
ity in rye. J Exp Bot 66:3275–3284
287
Williams JS (1962) The evaluation of a selection index. Biometrics
18:375–393
Windhausen VS, Atlin GN, Hickey JM, Crossa J, Jannink JL, Sor-
rells ME, Raman B, Cairns JE, Tarekegne A, Semagn K, Bey-
ene Y, Grudloyma P, Technow F, Riedelsheimer C, Melchinger
AE (2012) Effectiveness of genomic prediction of maize hybrid
performance in different breeding populations and environments.
G3 Genes| Genomes| Genetics 2:1427–1436
Yoshira T, Karasawa T, Nakatsuka K (2000) Yielding ability of winter
triticale in a heavy snow area of central Hokkaido. Crop Breed
Appl Biotechnol 2:213–218
Zhao Y, Gowda M, Liu W, Würschum T, Maurer HP, Longin FH,
Ranc N, Reif JC (2012) Accuracy of genomic selection in
European maize elite breeding populations. Theor Appl Genet
124:769–776
13