HAL Id: hal-01001607
https://hal.archives-ouvertes.fr/hal-01001607v2
Submitted on 19 Sep 2019
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entic research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diusion de documents
scientiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
High-throughput Phenotyping and Genomic Selection:
The Frontiers of Crop Breeding Converge
Llorenç Cabrera-Bosquet, José Crossa, Jarislav von Zitzewitz, María Dolors
Serret, José Luis Araus
To cite this version:
Llorenç Cabrera-Bosquet, José Crossa, Jarislav von Zitzewitz, María Dolors Serret, José Luis
Araus. High-throughput Phenotyping and Genomic Selection: The Frontiers of Crop Breeding
Converge. Journal of Integrative Plant Biology, Wiley, 2012, 54 (5), pp.312-320. �10.1111/j.1744-
7909.2012.01116.x�. �hal-01001607v2�
Journal of Integrative Plant Biology 2012, 54 (5): 312–320
Invited Expert Review
High-throughput Phenotyping and Genomic Selection:
The Frontiers of Crop Breeding Converge
F
Llorenc¸ Cabrera-Bosquet
1
, Jos
´
e Crossa
2
, Jarislav von Zitzewitz
3
,Mar
´
ıa Dolors Serret
4
and Jos
´
e Luis Araus
4
∗
1
French National Institute for Agricultural Research (INRA, UMR759), Ecophysiology Laboratory of Plants under Environmental Stress,
Montpellier F-34060, France
2
International Maize and Wheat Improvement Center (CIMMYT), El Bat
´
an, Texcoco CP 56130, Mexico
3
National Research Program on Rainfed Crops, National Institute for Agricultural Research, Est. Exp. La Estanzuela,
Colonia 70000, Uruguay
4
Unit of Plant Physiology, Faculty of Biology, University of Barcelona, Av. Diagonal, 643, Barcelona 08028, Spain
∗
Corresponding author
Tel: +34 93 402 1469; Fax: +34 93 411 2842; E-mail: jaraus@ub.edu
F
Articles can be viewed online without a subscription.
Available online on 16 March 2012 at www.jipb.net and www.wileyonlinelibrary.com/journal/jipb
doi: 10.1111/j.1744-7909.2012.01116.x
Jos
´
e Luis Araus
(Corresponding author)
Abstract
Genomic selection (GS) and high-throughput phenotyping have
recently been captivating the interest of the crop breeding com-
munity from both the public and private sectors world-wide. Both
approaches promise to revolutionize the prediction of complex
traits, including growth, yield and adaptation to stress. Whereas
high-throughput phenotyping may help to improve understanding
of crop physiology, most powerful techniques for high-throughput
field phenotyping are empirical rather than analytical and compa-
rable to genomic selection. Despite the fact that the two method-
ological approaches represent the extremes of what i s understood
as the breeding process (phenotype versus genome), they both
consider the targeted traits (e.g. grain yield, growth, phenology,
plant adaptation to stress) as a black box instead of dissecting
them as a set of secondary traits (i.e. physiological) putatively related to the target trait. Both GS and
high-throughput phenotyping have in common their empirical approach enabling breeders to use genome
profile or phenotype without understanding the underlying biology. This short review discusses the main
aspects of both approaches and focuses on the case of genomic selection of maize flowering traits and
near-infrared spectroscopy (NIRS) and plant spectral reflectance as high-throughput field phenotyping
methods for complex traits such as crop growth and yield.
Keywords:
Genomic selection; high-throughput phenotyping; NIRS; quantitative traits; SNPs.
Cabrera-Bosquet L, Crossa J, von Zitzewitz J, Serret MD, Araus JL (2012) High-throughput phenotyping and genomic selection: The frontiers
of crop breeding converge. J. Integr. Plant Biol. 54(5), 312–320.
Introduction
Grain yield, plant growth and stress adaptation are complex
traits controlled by many genes, usually with minor effects and
with a high occurrence of epistatic interactions (Li et al. 1997,
2001). This presents limited breeding advances for complex
traits. High-throughput phenotyping and genomic selection
(GS) of complex traits promise to revolutionize the breeding
process by accelerating generation-advance and improving the
precision of selection. On one hand, there is the fast, large-
scale evaluation of plant performance that aims to automatize
and standardize the phenotyping process. On the other hand,
C
2012 Institute of Botany, Chinese Academy of Sciences
High-throughput Phenotyping and Genomic Selection 313
there is the massive use of low-cost genotyping technologies
powered by advances in genomic sequencing and the advent
of single-nucleotide polymorphism (SNP) markers (Ingvarsson
and Street 2011). However, both methodological approaches
share in common their empirical nature and their strong de-
pendence on the advances in data gathering and processing
(together with the help of robotics). Thus, phenotyping is
evolving quickly from the concept of conventional breeding,
which relies on the direct measurement of the target trait
(e.g. yield) or even from analytical breeding (see Araus et al.
2008), which implies selecting the key secondary trait(s) (i.e.
other than the yield or the targeted trait itself), to the remote
inference of whole-plant growth, water status or even grain
yield using remote sensing approaches. For genotyping, classic
marker-assisted selection (MAS) that relies on the identification
of quantitative trait loci (QTL) for traits of interest has been
found to be far less successful than predicted two decades ago
when the target traits for selection were quantitative, such as
yield and adaptation to drought or other major abiotic stresses
(Maccaferri et al. 2008 and references herein). Phenotyping for
quantitative traits has been relatively ignored (even neglected)
until recently, when molecular biologists and breeders realized
that advances in molecular techniques may only be useful in
breeding for quantitative traits if they are based on reliable
phenotyping (Araus et al. 2003, 2008). In that context, and
taking advantage of the fast development in computing and
robotics, a burst of new technological approaches aimed at
shifting phenotying from an “art to a technology” has arrived
through the development of what are known as high-throughput
phenotyping platforms. That includes not just a large set of
commercially available instruments that allow this work t o
be done, but also the emergence of outsourcing services.
Thus, following the steps of genomic analyses and marker-
assisted selection, where outsourcing is a consolidated trend,
high-throughput phenotyping has been offered recently on a
server provider basis (www.phenofab.com; updated January
2012). The possibility to phenotype thousands of individuals,
faster, and with high precision, together with advances and
cost reduction in sequencing technologies (Davey et al. 2011),
allows GS to be applicable to plant breeding programs. GS
promises the power to predict individual genotype adaptability
to a specific environment (Crossa et al. 2010; Aguilar et al.
2011; Resende et al. 2012).
In the following discussion, the basics of high-throughput
phenotyping and genomic selection are briefly presented to-
gether with a few examples.
Genomic Selection: A Step Forward from
Marker-assisted Selection
The traditional approach of MAS, based on bi-parental or
association mapping panels, consists of detecting quantitative
trait loci (QTL) to further focus on significant regions for assisted
selection. Briefly, markers linked to genes are detected, the
most favorable alleles and their related effects are determined,
and then these are further validated. Nevertheless, MAS seems
effective only for major QTL effects (Araus et al. 2008; Macca-
ferri et al. 2008;vonZitzewitz et al. 2011), and not for complex
traits controlled by many genes with small effects, such as
grain yield and adaptation t o stresses. Further limitations in
MAS are the infrequent use of bi-parental populations for
germplasm improvement, and the statistical methods used for
mapping QTL, which rely on stringent thresholds and single
marker analysis models. On the other hand, genomic selection
(or genome wide selection) is an approach for improving
quantitative (i.e. complex) traits (Meuwissen et al. 2001) that
uses all the available molecular markers across the genome
and allows calculation of estimated genetic (breeding) values.
The following example may illustrate the need for empirical
models effective in capturing a large number of minor QTL.
In human association genetics, a set of 40 significant markers
explained only 5% of the variability in height while a set of
300 000 markers placed simultaneously in a model explained
up to 45% of the heritability for human height (Yang et al.
2010).
GS claims to act in a similar manner and intends to improve
the predictions of economically important traits in complex plant
and animal breeding programs. Instead of focusing on single
marker analysis models with the power to only detect relatively
large effects, GS is centralized in a genotypic characterization
of several markers (at low cost) to integrate simultaneously
(taking into account major and minor effects) in a predictive
model with novel statistical methods (de los Campos et al.
2009; Crossa et al. 2010; Vitezica et al. 2011). GS is based
on a predictive model that has been trained with a number of
individuals ("the training population") that reflects the diversity
of the breeding program being evaluated at the phenotypic
and genotypic level. As the number of markers increases in
the model, the prediction accuracies of genomic estimated
breeding values (GEBV) are expected to increase, whereas
the single marker effect is expected to decrease in absolute
magnitude. Once the model has proven successful, the plant
breeder is able to estimate the GEBVs for the next generation
cycle through the incorporation of only genomic data. These
GEBVs give an ideal criterion for selecting the best performing
lines.
Figure 1 shows a simple schematic diagram summariz-
ing the implementation of GS in a plant breeding program.
However, parametric regression models also provide the op-
portunity to examine marker effects and study the possible
differential response of markers in environments, that is, the
gene × environment interaction effect. In general, Bayesian
shrinkage methods do not have an associated test for detecting
chromosome regions; however, they can be routinely used for
QTL detection.
314 Journal of Integrative Plant Biology Vol. 54 No. 5 2012
Figure 1. Schematic diagram depicting the steps involved in
the use of genomic selection in a plant breeding program.
GEBVs, genomic estimated breeding values.
Although GS was first proposed about 10 years ago (Meuwis-
sen et al. 2001), reports on the use of GS in plants are few
and refer mainly to computer simulation studies such as the
research of Bernardo and Yu (2007), who concluded that GS
was superior to marker-assisted selection in maize. This delay
in the application of GS in plant breeding programs can be partly
explained by the high costs in genotyping germplasm with a
high density of markers at the time. It was not until recently that
relatively inexpensive marker technologies have become avail-
able as a service platform (i.e. http://www.diversityarrays.com;
updated January 2012). Further development of genotyping
methods based on direct sequencing of genomic digests have
reduced costs tremendously by relying solely on next genera-
tion sequencing and non-expensive array-based technologies
(Davey et al. 2011; Elshire et al. 2011). The advantage of
these methods is that full gain-of-sequence throughput is taken,
which is advantageous for polyploid species which cause
problems with hybridization, and polymorphism discovery is
simultaneous with genotyping. These last methods are one
step behind direct sequencing of whole genomes, which will be
applicable for GS once prices become affordable.
Diversity Array Technology (DArT) was recently used in
studies at the International Maize and Wheat Improvement
Center (CIMMYT), where de los Campos et al. (2009), Crossa
et al. (2010, 2011), P
´
erez et al. (2010) and Burgue
˜
no et al.
(2012) validated GS in plant breeding using genomic regression
and showed that models using molecular markers were more
accurate in predicting grain yield in wheat and maize than those
based on pedigree only. These were the first comprehensive
studies demonstrating that genomic selection could be useful
in plant breeding. These studies in maize and wheat based on
the multi-environment trials of CIMMYT indicated that models
for GS can have relatively high predictive ability for genetic
values of grain yield and other traits of economic interest under
contrasting environmental conditions. GS selection models with
pedigree and molecular marker information can be used effec-
tively for selecting individuals whose phenotypes for various
traits have yet to be observed under varying environmental
conditions, including for example severe water stress. All these
studies have indicated that the problem of model choice is
population and environment specific and that a ‘one-size-fits-
all’ approach to model choice in GS is not appropriate.
In addition to estimating genetic values, parametric mod-
els also provide information on ‘marker effects’ that can be
used to gain a better understanding of the important genomic
regions underlying the architecture of traits and genotype
× environment interaction. Principal component analysis of
estimated marker effects across environments provides a way
of identifying which markers contribute to positive genetic
correlations between environments, and which markers have
negative responses in other environments and therefore pro-
duce interaction with environments. In the following section, an
example is examined from the ongoing work of CIMMYT on GS
for predicting quantitative traits in plant breeding using dense
molecular markers and pedigree.
A Case Study: Patterns of Co-variability
of Estimated M arker Effects across
Environments for Maize Flowering
Genomic Data
Traits included here were female flowering (FFL) (or days to
silking), male fl owering (MFL) (or days to anthesis), as well as
the anthesis-silking interval (ASI) evaluated in 300 maize lines
under severe drought stress (SS) and in well-watered (WW) en-
vironments (Crossa et al. 2010). The first two component axes
of estimated effects of SNP markers in the six trait-environment
combinations (MFL-SS, MFL-WW, FFL-SS, FFL-WW, ASI-SS,
and ASI-WW) are depicted in
Figure 2. The correlation between
trait-environment combinations using marker effects and phe-
notypic data shows that some trait-environment combinations
are highly correlated (both phenotypically and genetically). The
pattern of correlations between estimated SNP effects reflects
the patterns of observed phenotypic correlations. Clearly, the
two groups of trait-environment combinations are dominated
more by the trait (ASI vs. FFL and MFL) and less by the en-
vironment (SS and WW). Phenotypic outcomes and estimates
of marker effects for ASI showed relatively small correlations
with those of FFL and MFL; this is because ASI is defined as
the difference between FFL and MFL, and these two traits are
positively correlated.
High-throughput Phenotyping and Genomic Selection 315
Figure 2. Biplot of the first and second principal component axes of the effects of 1 148 SNPs on 3 traits.
Biplot of the first and second principal component axes (Comp. 1 and Comp. 2) of the effects of the 1 148 SNPs on maize female flowering
(FFL), male flowering (MFL) and anthesis-to-silking interval (ASI) estimated from the full data model M-BL of the maize dataset in each
of two environments; severe water stress (SS) and well watered (WW). A total of six trait-environment combinations (FFL-SS, FFL-WW,
MFL-SS, MFL-WW, SS-ASI, and WW-ASI) were formed. Only the effects of the 19 SNPs that are located far from the center of the biplot
were identified with their corresponding SNP name (filled-in circles) (from Crossa et al. 2010).
Markers with relatively large (in absolute value) estimated
effects are identified by name in
Figure 2. The marker effects
on these traits should be interpreted differently than their effect
on grain yield, since the favorable marker allele decreases both
female and male flowering times, whereas for ASI, the optimal
marker should give an ASI of 0. The alleles whose estimated
effects are located in the left and upper left corner of the biplot
(i.e., PZA03551.1, PZA03578.1, PZA03222.1, PZA03385.1,
PZB01201.1, and PZB00118.2) increase FFL, MFL, and ASI
(they all have positive effects on all trait-environments combi-
nations), whereas those SNPs located on the opposite side of
the biplot (lower right corner) (i.e., PZA02587.16, PZA00236.7,
PZB0255.1, and PZA00676.2) decrease the value of FFL,
MFL, and ASI. Those SNPs whose presence is expected to
increase or decrease traits across environments can be viewed
as contributing to positive genetic correlations in FFL, MFL, and
ASI between environments.
Despite the high heritability (between 0.74 and 0.87) found
for flowering time and ASI in this maize trial, results show
substantial interaction between molecular marker effects and
environment. The biplot in
Figure 2 shows SNPs that had
very contrasting effects across environments. For example, the
minor alleles of SNPs whose estimated effects are located in
the upper right corner of the biplot (PZA03592.3, PZB01077.3,
and PZB02076.1) increase the anthesis-silking interval under
drought and well-watered conditions, but decrease days to
male and female flowering. In contrast, the minor alleles
of SNPs whose estimated effects are located in the oppo-
site quadrant of the biplot (lower left corner) (PZB00592.1,
PHM13183.12, and PZB01964.5) showed a complete rank
reversal with respect to the effects of SNPs PZA03592.3,
PZB01077.3, and PZB01077.3 on those trait-environment com-
binations, i.e., a decrease in ASI under SS and WW, and an
increase in male and female flowering times. These results
are suggestive of an important molecular marker effect ×
environment interaction, which in turn causes a genotype ×
environment interaction.
Genomic Traits with Genome ×
Environment
In a recent comprehensive study, Burgue
˜
no et al. (2012)
presented multi-environment (multi-trait) models for GS
and compared the predictive accuracy of these models
