MULTIPARENTAL POPULATIONS

Rapid Cycling Genomic Selection in a Multiparental

Tropical Maize Population
Xuecai Zhang,* Paulino Pérez-Rodríguez,† Juan Burgueño,‡ Michael Olsen,§ Edward Buckler,**
Gary Atlin,†† Boddupalli M. Prasanna,§ Mateo Vargas,‡‡ Félix San Vicente,*,1 and José Crossa‡,1
*Global Maize Program, International Maize and Wheat Improvement Center (CIMMYT), 06600 México D.F., México,
†Colegio de Postgraduados, CP 56230, Montecillos, 56230 México D.F., México, ‡Biometrics and Statistics Unit,
International Maize and Wheat Improvement Center (CIMMYT), 06600 México D.F., México, §International Maize and
Wheat Improvement Center (CIMMYT), Nairobi 1041-00621, Kenya, **United States Department of Agriculture,
Agricultural Research Service, Cornell University, Ithaca, New York 14853, ††Bill and Melinda Gates Foundation, Seattle,
Washington 98109, and ‡‡Universidad Autónoma Chapingo, 56230 Texcoco, México
ORCID ID: 0000-0001-9429-5855 (J.C.)

ABSTRACT Genomic selection (GS) increases genetic gain by reducing the length of the selection cycle, as KEYWORDS
has been exemplified in maize using rapid cycling recombination of biparental populations. However, no tropical maize
results of GS applied to maize multi-parental populations have been reported so far. This study is the first to multiparental
show realized genetic gains of rapid cycling genomic selection (RCGS) for four recombination cycles in a multi- population
parental tropical maize population. Eighteen elite tropical maize lines were intercrossed twice, and self- rapid cycling
pollinated once, to form the cycle 0 (C0) training population. A total of 1000 ear-to-row C0 families was recombination
genotyped with 955,690 genotyping-by-sequencing SNP markers; their testcrosses were phenotyped at four genomic
optimal locations in Mexico to form the training population. Individuals from families with the best plant types, selection
maturity, and grain yield were selected and intermated to form RCGS cycle 1 (C1). Predictions of the geno- realized genetic
typed individuals forming cycle C1 were made, and the best predicted grain yielders were selected as parents gains
of C2; this was repeated for more cycles (C2, C3, and C4), thereby achieving two cycles per year. Multi- genetic diversity
environment trials of individuals from populations C0, C1, C2, C3, and C4, together with four benchmark checks Multiparental
were evaluated at two locations in Mexico. Results indicated that realized grain yield from C1 to C4 reached Populations
0.225 ton ha21 per cycle, which is equivalent to 0.100 ton ha21 yr21 over a 4.5-yr breeding period from the MPP
initial cross to the last cycle. Compared with the original 18 parents used to form cycle 0 (C0), genetic diversity
narrowed only slightly during the last GS cycles (C3 and C4). Results indicate that, in tropical maize multi-
parental breeding populations, RCGS can be an effective breeding strategy for simultaneously conserving
genetic diversity and achieving high genetic gains in a short period of time.

In the last 20 yr, marker-assisted selection has been widely used in mic-assisted breeding (genomic selection, GS) incorporates all available
plant breeding where a few markers significantly associated with the marker information simultaneously into a model to predict the genetic
phenotypic trait are employed to predict the genetic value of the can- value of the candidates for selection (Meuwissen et al. 2001). In plants, a
didates for selection (Bernardo 2008, 2016). On the other hand, geno- computer simulation study (Bernardo and Yu 2007) showed that better
prediction accuracy of breeding and genetic values was achieved by
incorporating all markers, as compared to using a subset of markers
Copyright © 2017 Zhang et al.
significantly associated with QTL. This result was verified by Massman
doi: https://doi.org/10.1534/g3.117.043141
Manuscript received January 4, 2017; accepted for publication May 15, 2017; et al. (2013), who used a biparental temperate maize population derived
published Early Online May 22, 2017. from a cross between two distinct heterotic groups (B73 and Mo17); the
This is an open-access article distributed under the terms of the Creative testcrosses were evaluated under well-watered conditions, and the pop-
Commons Attribution 4.0 International License (http://creativecommons.org/ ulation advanced using rapid cycling GS (RCGS where all markers are
licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction
in any medium, provided the original work is properly cited.
used for prediction) and marker-assisted recurrent selection (MARS
1
Corresponding authors: CIMMYT, Apdo. Postal 6-641, 06600 México D.F., where only significant markers are used for prediction). Massman et al.
México. E-mail: [email protected]; and [email protected] (2013) reported that RCGS had a superior response for stover yield, as

Volume 7 | July 2017 | 2315

well as stover and grain yield indices that were 14–50% higher than under selection and association mapping. In a recent article, Hoffstetter
those of MARS. et al. (2016) used 47 bi-parental crosses, including 23 parental wheat
In tropical maize, Beyene et al. (2015) evaluated realized genetic lines as a training set, and 17 half-sib lines as a validation set, with some
gains in grain yield from RCGS in eight biparental maize populations in of the lines from the training set included in the testing set. Prediction
drought stress environments. The authors found that the average gain accuracy using subsets of the training population to predict the valida-
from RCGS per cycle across eight populations was 0.086 ton ha21 and tion sets ranged from 0 to 0.85. In maize, Lehermeier et al. (2014)
that hybrids derived from cycle 3 produced 7.3% (0.176 ton ha21) evaluated prediction accuracy in 21 biparental doubled haploid popu-
higher grain yield than those developed through the conventional ped- lations, including 10 dent kernel type related populations, and 11 flint
igree breeding method. RCGS in biparental populations offered the kernel type related populations. The authors found that prediction
advantage of significant time efficiency over conventional breeding combining several half-sibs gave similar or higher prediction accuracy
methods, as up to three cycles of RCGS can be conducted within a than predictions within biparental populations.
year. Interestingly, Beyene et al. (2015) pointed out that the average However, several theoretical complexities and challenges arise when
genetic gain per year in tropical maize grain yield using RCGS was three attempting to perform GS in MPP. Plant breeders usually work with sets
times higher than that achieved by using conventional pedigree-based of full-sib families generated from crosses of inbred parents that vary in
phenotypic selection in drought stress environments. Furthermore, size so that extensive LD exists within each family; however, different LD
Vivek et al. (2016) recently reported a study on realized genetic gains patterns must exist across families. Therefore, because, in a typical MPP,
in grain yield using two biparental populations generated by crossing the breeder is faced with LD across a large set of families, the level of LD as
two elite Asian maize inbred lines with an African drought tolerant line. well as the marker density are important factors to be considered when
Cycle 1 (C1) was formed by recombining the top 10% of the F2:3 applying GS to MPP. Zhong et al. (2009) using marker data on 42 two-
families. Cycle 2 (C2) was derived using two different methods: (1) row spring barley inbred lines simulated MPP with high and low LD
phenotypic selection (C2-PS); and (2) GS (C2-GS). Results showed populations generated from multiple inbred crosses. The authors sug-
that C 2 -GS top-crosses produced 4–43% higher grain yield than gested a trade-off between the model-method’s ability to capture LD
C2-PS top-crosses. between marker and QTL vs. its ability to exploit marker-based re-
For RCGS within biparental populations, prediction accuracy is latedness of individuals. Genomic relationship information (Best Lin-
achieved thanks to high linkage disequilibrium (LD), no pedigree, ear Unbiased Predictor, GBLUP) is more valuable than LD information
and no family substructure (Crossa et al. 2014; Zhang et al. 2015). given by models that use marker effects. However, when markers were
However, predictions across biparental populations will be poor if un- in strong LD with QTL with large effects, models based on marker
related biparental populations with different allelic diversity are used as effects were better predictors than models based on the genomic re-
the training population. Furthermore, Jannink et al. (2010) outlined lationship between individuals.
some of the disadvantages of using GS applications in biparental pop- The reliability of GS is greatly influenced by the number of pheno-
ulations: (1) separate model training is required for each biparental types; therefore, combining data sets from MPP should increase the
cross; GS should be applied to the entire population; and (2) the first reliability of GS by making it more efficient and attractive for use in
generation of progeny from a cross needs to be phenotyped and can- breeding. However, when combining populations, allele frequencies, LD
didates cannot be selected on the basis of prior information, a practice and segregating haplotypes are different in different populations. Thus,
that slows down the breeding cycle. Biparental populations have also when the marker effects are different between the different combined
been widely used for detecting and mapping QTL. However, QTL populations, this can reduce the reliability of GS. de Roos et al. (2009)
mapping power and resolution might be comparatively reduced in bi- studied this problem by combining two simulated cattle populations
parental mapping populations as the number of segregating causal loci that diverged for certain number of generations; the authors found that
is very low, because large blocks of parental chromosomes are preserved increased genomic accuracy is achieved when all populations are com-
(Cavanagh et al. 2008; Thepot et al. 2015; Lehermeier et al. 2014). The bined in one training population but increasing marker density is re-
problem of limited allelic diversity in one genetic background which quired when the diversity of the combined population increases.
occurs in biparental populations can be overcome by the use of multi- Despite the above theoretical and simulation results, which show that
parental populations (MPP) with greater allelic diversity and from genomic-enabled prediction accuracy of MPPs is higher than the
different genetic backgrounds (Verhoeven et al. 2006), along with in- accuracy achieved within a single population, studies implement-
creased polymorphism and recombination as compared to biparental ing RCGS in MPPs have not been reported. The research presented
populations (Ahfock et al. 2014). in this paper was initially conceived in 2009 and started in 2010. It
Different MPPs have been constructed with the aim of increasing included an initial MPP made up of 18 elite tropical maize lines
precision in fine mapping. For example, researchers have combined intercrossed twice and self-pollinated once to form the cycle 0 (C 0 )
different biparental populations using factorial crosses, partial or com- training population. A total of 1000 ear-to-row C0 families were
plete diallels (Rebai and Goffinet 1993), or circular crosses (Cavanagh genotyped with dense GBS markers, and their testcrosses were
et al. 2008; Huang et al. 2012). The designs and analyses of these types phenotyped at four locations in Mexico to develop genomic pre-
of populations have been extended to genomic-enabled studies using diction models. One cycle of phenotypic selection (C 0–C 1) and
different schemes and designs for efficient training, and for testing three cycles of RCGS (C1 –C4) were carried out. The main objec-
MPPs in different species. tives of this study were: (1) to report the realized genetic gains of
Multi-parental populations are useful for mapping extensive num- four cycles (C 1, C2, C3, and C4), plus the original training pop-
bers of loci. Thepot et al. (2015) used a multi-parental mapping ulation (C0) in multi-environmental field trials of RCGS-assisted
population created after 12 generations of recombination among breeding evaluated together with four benchmark checks in two
60 founder wheat lines. The 12 generations of recombination broke Mexican environments (locations), and (2) to investigate the ge-
up the LD and the existing population structure of the original pop- netic diversity of the families within each RCGS selection cycle to
ulations. This approach to fine mapping helped to identify 26 genomic assess the level of genetic diversity after three cycles of rapid
regions, six of which carried flowering QTL, and allowed detecting loci cycling GS.

2316 | X. Zhang et al.

Figure 1 Breeding scheme used in the MPPs reported in this study.

MATERIALS AND METHODS (CML495/CML549) from the complementary heterotic group (dent
type kernel) was used to generate testcrosses. All pollination activities
Developing the training population from 18 tropical
were conducted at CIMMYT’s experiment station in Agua Fria, Puebla.
maize inbred lines
The training population (C0) for developing genomic prediction
The RCGS experiment was designed in 2009 as part of the MasAgro
models was formed with the best 1000 selected S2s. Testcrosses of all
project funded by Mexico’s Secretariat of Agriculture, Livestock, Rural
1000 selected S2 were planted using a partial replicated design with 25%
Development, Fisheries and Food (SAGARPA, its acronym in Spanish)
of replicated genotypes at four optimal Mexican locations. Phenotypic
through the Sustainable Modernization of Traditional Agriculture pro-
data were collected at all locations for .10 agronomic traits, including
gram (MasAgro; http://masagro.mx).
grain yield at 12.5% moisture content (GY), anthesis date (AD), silking
The steps in the breeding scheme used for RCGS are shown in
date (SD), plant height (PH), ear height (EH), and moisture content
Figure 1. In total, 18 CIMMYT tropical maize inbred lines (CML247,
(MOI). For each S2 family, DNA was extracted by bulking equal
CML264, CML448, CML494, CML498, CML531, CLRCW72, CLRCW75,
amounts of leaf tissue from 15 individual plants. Genotyping-by-
CLRCW76, CLRCW93, CLRCW100, CLRCW260, CLWN201, CLWN228,
sequencing was performed at Cornell University Biotechnology Re-
CLWN229, CLWN247, CLG2312, and CLSPLW04), widely used in
source Center as described by Wu et al. (2016), where 955,690 SNPs
lowland tropical breeding environments, were crossed as parents to
were generated for each DNA sample. In the training population, the
form the training population through twice intermated pollination
genomic prediction model was developed by using only 331,740 filtered
and one self-pollination; the parents were selected based on their gen-
SNPs with minor allele frequency (.0.05), and where the missing data
eral combining ability for grain yield and per se, visual evaluation in-
rate was ,10%.
formation for major stress tolerance and disease resistance in lowland
tropical breeding environments. All 18 original parents tended to group
in heterotic pattern group “B” (flint type kernel). Cycle 0 (C0) phenotypic selection and formation of
In the 2010B season, half-diallel crosses were made between the cycle 1 (C1)
18 original parents to generate all possible F1 progenies (Figure 1). In the In C0, phenotypic selection was conducted by ranking the grain yield of
2011A season, all F1 were planted ear-to-row and intermated to form the 1000 S2 testcrosses. The best 50 respective S2 families were selected
the S1 population; then, all the F1 were separated into two groups of and planted ear-to-row, 25 plants per family (Table 1). Cycle 1 (C1) was
equal size. Bulk pollen from the first group was used to pollinate all formed by intermating the 50 selected S2 families. The 50 families were
plants of the other group and vice versa; three ears were harvested from divided into two equal groups, and bulk pollen from the first group was
each F1 family, and equal amounts of seed from each selected F1 ear used to pollinate all plants in the other group and vice versa. Based on
were bulked to form the subsequent generation for planting. In the visual evaluation of flowering time, plant type, plant/ear height, well-
2011B season, 4800 S1 individuals were planted and self-pollinated filled ears, and reaction to naturally occurring major diseases, along with
and advanced to S2. The best 1000 S2 ears were selected, and planted among-family and within-family selection, 157 ears (1–6 ears from each
ear-to-row in the 2012A season (Table 1). A single-cross tester selected family) were harvested and shelled individually to form C1.

Volume 7 July 2017 | Genomic Selection in Multi-Parental Population | 2317

n Table 1 Number of families and individual plants sown, Phenotypic evaluation of the selection cycles for
selected, and advanced in each breeding cycle and among-family, assessing genetic gains—benchmark checks and
within-family, and total selection intensity experimental designs in multi-environment trials
Cycle A total of 233 testcrosses was tested in the field to estimate the realized
C0 C1 C2 C3 C4 genetic gains; these entries belong to different groups. One group of
entries (48) represents selection cycle C0, which is a subset of the best
No. of families sown 1000 157 91 44 –
No. of families selected 50 25 18 22 – 50 families selected from the training population. Another group of
No. of plants sown per 25 25 25 25 – entries represents RCGS cycles C1 (47 entries), C2 (48 entries), C3
family (43 entries), and C4 (43 entries), and the last group of entries comprises
No. of ears selected 157 91 44 22 45 the four benchmark checks (two local checks, one commercial check,
Among-family selection 5% 16% 20% 50% – and one experimental baseline check formed by testcrossing all 18 orig-
intensity inal parents with a single-cross tester (CML495/CML549). All the en-
Within-family selection 12.60% 2.30% 1.90% 2.00% – tries were crossed with a single-cross tester (CML495/CML549). Their
intensity
testcrosses were planted at two locations in Mexico (Agua Fria and
Total selection intensity 0.60% 0.40% 0.40% 1.00% –
Tlaltizapan) in a modified split-plot design where the selection cycles
were the main plots and the entries within each selection cycle were the
subplots. The experimental design within each selection cycle (main
Rapid cycling recombination of GS cycle 1 (C1), cycle plot) was an alpha-lattice design with two replications per location. The
2 (C2), and cycle 3 (C3)
four benchmark checks were repeated in each subplot and planted
In C1, 157 selected ears (Table 1) were planted ear-to-row, 25 plants together with the entries belonging to the different GS cycle; for exam-
per family. DNA was extracted from the bulk tissue and shipped to ple, the four checks were planted in the two replicates of the subplot
Cornell University Biotechnology Resource Center for genotyping- where entries from cycle C0 were planted.
by-sequencing. The Genomic Best Linear Unbiased Predictor
(GBLUP) model (VanRaden 2007, 2008) implemented in the BGLR Statistical analyses
package was used for genomic prediction. Genomic estimated Phenotypic data were collected at the two locations for the main
breeding values were calculated for all 157 C1 families; among-family agronomic traits including GY, AD, SD, PH, EH, and MOI. A linear
selection was implemented based on genomic estimated breeding value mixed model was fitted to the data considering the incomplete block
information. The top 25 families were selected and intermated to form within replicate as random effects, and locations, cycles (main plot),
the C2 population. The 25 families selected were divided into two entry within cycle (subplot), cycle · location interaction, and entry
equal groups. Bulk pollen from one group was used to pollinate all within cycle · location interactions as fixed effects. Random errors
plants in the other group and vice versa. Within-family selection was are assumed to be identically and independently normally distributed
implemented based on visual evaluation of flowering time, plant type, with mean zero and homogeneous variance. ANOVA for GY were
plant/ear height, well-filled ears, and reaction to naturally occurring performed including the evaluated entries from the selection cycles
major diseases. A total of 91 ears were harvested, and shelled indi- and the checks. Genetic gain response was assessed by regressing mean
vidually to form C2. In C2 and C3, the recombination protocol was GY values on the selection cycle means (C0, C1, C2, C3, and C4) within
repeated. The number of families and individual plants planted per each location and combined across both locations. ANOVA for AD,
cycle; number of families selected for next cycle recombination; num- SD, PH, EH, and MOI were performed for each location and combined
ber of ears harvested per cycle; and selection intensity information are across both locations.
listed in Table 1. After C3 recombination, 45 ears were harvested and
shelled individually to form C4; within-family selection was imple- SNP genotyping
mented based on visual evaluation of agronomic traits. Families from For each cycle, bulk DNA of each planted family was sent to the
Cycle 4 (C4) were not genotyped, as this was the last rapid cycling Biotechnology Resource Center of Cornell University for genotyping-
recombination. by-sequencing. The number of DNA samples used for genotyping were

n Table 2 Mean of GY (ton ha21) for each genomic cycle C0, C1, C2, C3, and C4, broad-sense heritability (H2), and mean of the four testers
at each location (Agua Fria and Tlaltizapan), and combined across the two locations
Agua Fria Tlaltizapan Combined
Cycle Entry H2 Checks Entry H2 Checks GY H2 Checks
C0 6.65 0.27 5.47 10.40 0.65 8.08 8.52 0.42 6.77
C1 6.49 0.06 5.73 10.29 0.59 9.32 8.40 0.63 7.52
C2 7.02 0.26 6.02 10.20 0.46 9.30 8.62 0.47 7.52
C3 6.88 0.38 5.64 10.95 0.59 9.31 8.92 0.67 7.52
C4 7.13 0.21 5.70 10.96 0.25 9.30 9.05 0.43 7.61
LSD0.05 (C0–C4) 0.402 — — 0.412 — 0.252 —
LSD0.05 (C1–C4) 0.408 — — 0.404 – 0.191 —
Average gain per cycle (C0–C4) 0.131 — — 0.177 — 0.158 —
Average gain per cycle (C1–C4) 0.171 — — 0.276 — 0.225 —
The average genetic gain in GY across cycles was estimated for each location and across both locations including all selection cycles (C0–C4), and including only the
genomic selection cycles (C1–C4). Least significant differences (LSD) test at the 0.05 probability level including all selection cycles (C0–C4) and only the genomic
selection cycles (C1–C4). The highest value is indicated in boldface.

2318 | X. Zhang et al.

n Table 3 Means of entry and checks for traits anthesis days (AD, days), silking days (SD, days), plant height (PH, centimeter), ear height
(EH, centimeter), and moisture content (MOI, %) in each cycle across the two locations
AD SD PH EH MOI
Cycle Entry Check Entry Check Entry Check Entry Check Entry Check
C0 56.54 56.58 57.22 57.62 250.86 249.86 130.71 136.84 17.06 17.64
C1 56.54 57.02 57.35 58.11 250.23 247.39 129.55 131.77 18.61 19.45
C2 56.60 56.82 57.05 57.71 254.22 251.15 130.86 135.54 18.51 19.03
C3 56.11 56.27 56.80 57.31 258.11 252.05 133.00 134.18 19.42 19.88
C4 56.49 56.51 57.26 57.75 260.17 255.33 136.80 137.03 17.35 17.40
LSD0.05 (C0–C4) 0.210 — 0.233 — 2.882 — 2.408 — 0.254 —
LSD0.05 (C1–C4) 0.167 — 0.173 — 2.205 — 2.306 — 0.230 —
Least significant differences at the 0.05 probability level including all cycles (LSD0.05 (C0–C4)) and including only the genomic selection cycle (LSD0.05 (C1–C4)).

1000, 157, 91 and 44 in C0, C1, C2 and C3, respectively. Families from C4 HiSeq2000 (Elshire et al. 2011). SNP calling was performed using the
were not genotyped. Genotyping-by-sequencing, SNP calling, imputa- TASSEL GBS Pipeline, and a GBS 2.7 TOPM (tags on physical map) file
tion and filtering were performed as described by Wu et al. (2016). was used to anchor reads to the Maize B73 RefGen_v2 reference
Briefly, genomic DNA was digested with the restriction enzyme ApeKI. genome (Glaubitz et al. 2014). Imputation was carried out with the
GBS libraries were constructed in 96-plex and sequenced on Illumina FILLIN method in TASSEL 5.0 (Swarts et al. 2014), which anonymized

Figure 2 Distribution of parents, cycle C0 entries (A) and the selected parents, and cycle C1 entries (B) and the selected parents based on rapid
cycling genomic selection-assisted recombination.

Volume 7 July 2017 | Genomic Selection in Multi-Parental Population | 2319

Figure 3 Distribution of parents, cycle C2 entries (A), cycle C3 entries (B) and the selected parents based on rapid cycling genomic selection-
assisted recombination.

GBS 2.7 haplotypes made from 8000-site windows. In total, 955,690 RESULTS
SNPs were generated for each sample, filtering was performed
Heritability and prediction accuracy of GY in the
with minor allele frequency (.0.05) and the missing data rate
training population
was ,10%.
The combined GY heritability across both locations was 0.34, while GY
heritability at individual locations was 0.48, and 0.19 in Agua Fria and
Assessing the genetic diversity of the selection cycles
Tlaltizapan, respectively. Low-to-intermediate GY heritability was ob-
Based on genomic data, we computed two genetic diversity indices
served in the individual location analysis and combined analysis, mainly
between the families of the different selection cycles as well as the
because GY is a complex trait. To mimic future prediction problems we
18 parents. We calculated Pthe Shannon Diversity Index of the sample for will face, we implemented a fivefold random cross-validation with
each selection cycle as A1 Aa¼1 ^pa lnð^pa Þ; where ^pa is the frequency of the
100 replicates using entries in C0 (training population) for GY; the
major allele in the ath marker over the entire sample, and A is the total
mean correlation between the predicted and observed values was 0.55.
number of markers. The expected proportion of heterozygous loci per
individual
Pwas computedP i as2 the mean of heterozygosity for each marker
as 0 # L1 Ll¼1 ð1 2 na¼1 ^pla Þ # 1, where ^pla is the frequency of the Realized genetic gains from rapid cycling recombination
major allele in the a th marker of the l th individual, and L is the number of GS for grain yield
of individuals. A total of five groups of entries from C0, C1, C2, C3, and C4 plus four
Multidimensional scaling (MDS) was performed with the TASSEL checks were used for field evaluation at two Mexican locations (Agua
software (http://www.maizegenetics.net/tassel) to assess the genetic Fria and Tlaltizapan). Mean grain yield for each cycle and average gains
similarity of all the materials in each selection cycle. per cycle are shown in Table 2. The Tlaltizapan location had the highest
mean yield, with C4 reaching 10.96 ton ha21. At both individual lo-
Data availability cations, the average performance across all C4 entries surpassed the
The phenotypic and genotypic data for the training population (cycle C0) grain yield performance of the other cycles; average grain yield perfor-
evaluated in four sites, the phenotypic and genotypic data for the eval- mance across all C4 entries was 7.13 ton ha21, and 10.96 ton ha21
uation of the entries from the different selection cycles (C0, C1, C2, C3, for Agua Fria and Tlaltizapan, respectively. Also, the combined analyses
and C4), as well as a brief GUIDE can be found in the link http://hdl. of the two locations showed an increase in mean grain yield in C4 of
handle.net/11529/10927. A marker information file and characteristics 9.05 ton ha21 over that achieved in C3 (8.92 ton ha21) and over the
of the genetic materials are also included in the link. other GS cycles.

2320 | X. Zhang et al.

 selection cycles. Mean performance of the checks in selection cycle C0
was consistently lower than the mean performance in the other selec-
tion cycles.
The average gains per cycle for each location and combined across
the two locations ranged from 0.131 ton ha21 to 0.177 ton ha21
when considering all cycles (C0–C4), and from 0.171 ton ha21 to
0.276 ton ha21 when considering only rapid cycling GS (C1–C4).
For the combined location analyses, the realized genetic gains were
0.158 ton ha21 and 0.225 ton ha21 for cycles (C0–C4) and cycles
(C1–C4), respectively. The realized genetic gains due to RCGS were
highest in C4 at the two locations and combined across them.

Changes in the mean for unselected flowering,

moisture, and height traits
The effects of genomic selection on unselected flowering, moisture and
height traits are shown in Table 3. On average, anthesis and silking days
of the entries representing C4 did not increase with respect to their
averages in early GS cycles (C0, C1, C2, and C3). They ranged from 56 d
for anthesis and 57 d for silking for all genomic selection cycles, and
Figure 4 Plot of the two dimensions of the multi-dimensional scaling showed good general synchrony between both flowering times. How-
analysis including all samples of the original parents and families in ever, GS produced taller plants and ear insertions during cycles C3 and
each cycle.
C4 than during cycles C1 and C2. Grain moisture content did not seem
to have been greatly affected after the three cycles of RCGS.
For each location and combined, the entries representing the first The plant and ear heights for the two latest GS cycles were 6–10 cm
selection cycle (C1) had lower GY than entries representing the base higher than the plant and ear heights of the maize plants for the two
selection cycle (C0); this was due to the fact that the parents of the C1 earlier GS cycles (Table 3). The phenotypic variance (data not shown)
population (Table 1) were not selected based on the grain yield perfor- of the entries for each GS cycle varied during the latest cycles (C3 and
mance of the testcrosses per se; instead, among- and within-family C4), and tended to decline with respect to the early cycles. Similarly,
selection was conducted based on visual evaluation for flowering time, trends in broad-sense heritability are smaller in the last cycle (C4) than
plant type, plant/ear height, well-filled ears, and reaction to naturally in early cycles (C0 and C1). The mean of the entries within each selec-
occurring major diseases. The other important reason is that the best tion cycle, and the mean of the benchmark checks did not differ much
50 selected families were used to represent selection cycle C0 in the for the different traits. For example, the check had 56 d to anthesis
genetic gain evaluation study (rather than the random selected fami- and 57 d to silking. In general, the checks had smaller plant height,
lies). In C3, GY substantially increased in TL and combined across two taller ear height and similar grain moisture than the entries in the
locations with respect to previous selection cycles. For analyses in Agua selection cycle.
Fria location, GY declined slightly to 6.88 ton ha21 for C3 (Table 2). The phenotypic correlation between the different traits and grain
However, in the subsequent genomic selection cycles, the increases in yield changed from different selection cycles (data not shown). Results
realized genetic gains are clear for each location and combined across indicated positive correlations between grain yield and days to silk
locations. ranging from 0.3 to 0.5 for the different selection cycles. Negligible and
All selection cycles had higher average grain yield than the corre- negative correlations between grain yield and plant height, anthesis days,
sponding mean of the checks for individual locations and combined. and moisture content were found.
Concerning broad-sense heritability, there was a decline in cycles C1 and
C4 as compared to the base cycle (C0) in Agua Fria and Tlaltizapan; Genetic diversity of the rapid cycle recombination of GS
however, the opposite occurred when combining cycles C1 and C3 in The diversity structure pattern including the 18 parents, the C0 families
both locations as compared to the heritability in the base cycle (C0). In and the individuals selected as parents of C1 is displayed in Figure 2A;
the combined analyses across both locations, broad-sense heritability the selected individuals are well spread along the three dimensions,
(H2) showed no decrease. In general, results showed that at each loca- and should capture most of the diversity in the C0 families. The
tion and combined, while there was a decrease in GY from C0 to C1 original parents, the C1 families and the selected individuals (that
there were also important increases in realized genetic gains for trait GY form the parents of C2 selection based on genomic prediction) are
at the two locations from C2 to C3 and from C3 to C4. The mean grain depicted in Figure 2B. The C1 families are located between dimen-
yield of the benchmark checks varied slightly for each cycle at Agua Fria sions 1 and 3, close to four of the original parents located in this
but stayed fairly constant at Tlaltizapan and combined locations for all region of the figure.

n Table 4 The Shannon Diversity Index, heterozygosity, and number of SNPS of the 18 original parents, the number of families in cycles
C0–C3 (in parentheses), and the selected parents in C0–C3, and including all the entries
Parents C0 (1000) C0 (50) C1 (157) C1 (25) C2 (91) C2 (18) C3 (44) C3 (22) All Entries
Shannon’s Index 0.0661 0.0728 0.020 0.0776 0.052 0.0765 0.043 0.0588 0.063 0.0740
Heterozygosity 0.1104 0.1226 0.1208 0.1297 0.1250 0.1276 0.1228 0.0973 0.0923 0.1245
Number of SNP markers 950,248 952,825 943,344 951,390 947868 953,199 953,453 954,058 954,924 954,960
Numbers in parentheses refer to the size of the cycle population and the selected parents to form the subsequent cycle.

Volume 7 July 2017 | Genomic Selection in Multi-Parental Population | 2321

 Diversity Index and the heterozygosity for cycles C0–C4, along with the
selected parents from C0 to C4 and all other entries, are displayed in the
barplots in Figure 5. Genetic diversity did not decline in the initial
cycles and the trends in Shannon’s Diversity Index and heterozygosity
even showed a slight increase in cycles C1 and C2. However, there was a
decrease in the three diversity measurements for the population repre-
sented by the C3 entries, as well as for the parents selected to form C4.
Note that entries from C4 were not included because they were not
genotyped.

Changes in the status of marker frequency

For each SNP marker, we measured the change in the allele frequency of
the 18 parents’ original parents that formed C0, and the allele frequency
of the C3 and C4 entries. Initially, allele frequency was calculated for the
18 parents, and for the entries of cycles C3 and C4; SNP markers with
allele frequencies between 0.05 and 0.95 were considered polymorphic
markers, whereas markers with allele frequencies ,0.05 or .0.95 were
considered monomorphic markers. Using these criteria, we found that
Figure 5 Bar-plot of the Shannon Diversity Index (blue) and hetero- from a total of 1120 SNP markers (0.117% of all markers) that changed
zygosity (brown and light pink) for the original parents, individuals from their polymorphic/monomorphic status, 123 markers swapped the ma-
cycles C0–C3 and the selected entries in light blue and light pink. jor allele frequency, 968 markers became monomorphic in C3–C4 al-
though they were polymorphic in the 18 parents and C0, and only
The original 18 parents, the C2 families, and the parents selected to 29 markers were polymorphic in cycles C3 and C4, although they were
form the next selection cycle are shown in Figure 3A. The C2 families monomorphic in the 18 parents and cycle C0 (Table 5). Chromo-
and the selected parents are located between dimensions 1 and 3, somes 1–4 had a low percent change in allele frequency (0.080–
clearly heading in the direction of two of the original parents located 0.097%), whereas the rates of change in chromosomes 5 and 6 were
in this region of the three-dimensional figure. Finally, Figure 3B in- 0.177, and 0.139%, respectively. Chromosomes 7–10 had a percent
cludes the original 18 parents, the C3 families and the selected parents change in allele frequency ranging from 0.118 to 0.131%. Almost two
that were intermated in RCGS to form C4. Clearly, the C3 families and thirds of the markers (614,663 markers, 64.37%) changed their allele
the selected parents that form C4 are concentrated around the two frequencies without becoming monomorphic and 33.99% (324,575) of
original parents located in the upper region of the figure, between the markers changed their allele frequency by ,15%.
dimensions 1 and 3. However, a direct comparison between the genetic Markers with changes in their frequency were clustered accord-
diversity of different populations (C0–C4) may be confounded by the ing to the physical distance in the map; SNPs with a physical
differences in population size, and also by the different levels of in- distance ,1000 bp were considered a cluster (or haplotype). A total
breeding in the different selection cycles. of 88 clusters (haplotypes) of different sizes were found with three or
Figure 4 depicts the two plot dimensions of the multidimensional more SNPs that changed their frequency. The distribution and size
scaling with the 18 parents and all families in selection cycles C0–C4. of the clusters as well the type of change are shown in Table 6. For
The C3 families and the selected parents that form C4 are located to- example, chromosome 1 had one cluster with four markers that
ward the upper left quadrant of the biplot in the same direction as one changed their frequency; this indicated that there were at most
of the original parents. 4000 bp units where these four markers were located. Most of the
Concerning the genetic diversity of the different C0–C4 entries, Table changes correspond to SNP markers that became monomorphic in
4 shows the values of Shannon’s Diversity Index, the heterozygosity and cycles C3 and C4 although they were polymorphic in the 18 parents
the number of SNPs for the 18 parents, cycles C0–C4, the selected and cycle C0. Most of the clusters were found on chromosome 5
parents from C0 to C4, along with all the entries. Results of Shannon’s (Table 6).

n Table 5 Number of SNP markers with allele swaps, number of polymorphic markers that became monomorphic and number of markers
that were monomorphic and became polymorphic from [parents-C0] to [C3–C4]
Number of Number of Polymorphic Number of Monomorphic Total SNPs with Changes Total Number
Chromosome Allele Swaps to Monomorphic to Polymorphic N % of SNPs

1 12 124 7 143 0.096 148,745

2 12 105 1 118 0.102 115,152
3 7 97 1 105 0.097 108,195
4 11 64 1 76 0.080 94,716
5 20 166 9 195 0.177 110,303
6 12 91 3 106 0.139 76,450
7 7 90 2 99 0.123 80,514
8 10 86 96 0.118 81,427
9 21 71 2 94 0.130 72,355
10 11 74 3 88 0.131 67,103
Total 123 968 29 1120 0.117 954,960

2322 | X. Zhang et al.

n Table 6 Distribution and size of clusters of SNPs with changes in their polymorphic status by chromosome
Allele Swap Polymorphic to Monomorphic Monomorphic to Polymorphic
Number of SNPs Number of SNPs Number of SNPs
Chromosome 3 4 5 7 3 4 5 6 7 8 9 3 4 5 Total

1 1 3 4 2 1 11
2 1 1 4 1 1 1 9
3 1 1 2
4 2 2 1 1 6
5 2 9 3 2 1 1 1 19
6 1 3 2 1 1 8
7 2 2 1 1 6
8 2 4 2 2 10
9 1 2 1 3 1 8
10 1 4 3 1 9
Total 6 1 1 1 31 21 12 5 4 2 1 1 1 1 88

Figure 6 depicts the location of the clusters for each chromosome in etc. This also helped to broaden and maintain the genetic diversity
the genome. For example, chromosome 1 has one cluster with four observed in C1 and C2, which later declined in C3. As already men-
markers that changed their frequency (green color), one cluster of four tioned, the other reason of lower GY observed in C1 compared to C0 is
markers that changed from monomorphic in the 18 parents and cycle that the best 50 selected families (not the random selected families)
C0 to polymorphic in cycles C3 and C4 (blue color), and three clusters were used to represent selection cycle C0 in the genetic gain evaluation
with three markers, four clusters with four markers, and two clusters study. The GY mean of the best 50 selected families is much higher than
with six markers that changed from polymorphic in the 18 parents and that value of the 50 random selected families.
cycle C0 to monomorphic in cycles C3 and C4 (black color). As shown In terms of the prediction models used to predict the genetic values of
in Table 6, most of the clusters with changes in their allele frequency the entries to be selected in each genomic cycle, we used the direct
occurred in chromosome 5. genotyping-by-sequencing marker as biallelic. Since a multi-parental
(not a biparental) population was used, haplotype rather than biallelic
DISCUSSION marker could have been used in order to attempt to capture the whole
Previous studies on temperate and tropical maize showed realized gains allelic diversity. However, the problem on how to define the length of the
of RCGS in biparental populations (Massman et al. 2013; Beyene et al. haplotype segment in each chromosome could impose a major draw-
2015; Vivek et al. 2016). In this study, our results showed realized gains back for using this approach; different haplotypes methods exist but
of RCGS in a multi-parental tropical maize population that originated none of them seems to give clear superiority in terms of genomic-enabled
from crosses of 18 CIMMYT elite tropical maize lines. From a practical prediction accuracy.
breeding perspective, multi-parental populations might not be an at-
tractive option because the mean of 18 parental lines might be lower Realized genetic gains per unit of time
than the mean of the best few lines that could be used in biparental To compute the realized genetic gains per year (ton ha21 yr21), it is
crosses; however, as diversity becomes an important issue in GS, multi- necessary to account for the number of cycles per year (two cycles per
parental populations offer the opportunity to maintain diversity, while year in this study), and also for the time from the initial cross to the last
still achieving rapid cycles with high realized grain yield genetic gains cycle (4.5 yr from F1 development to harvesting the C4 in this study).
achieved in a shorter period of time, as found in this study. As for the
decrease in genetic diversity, this is not of much concern in a short-term
selection (four to five cycles), especially if the new developed lines from
C4 are crossed with other lines for further breeding.
Trends in the realized genetic gains of multi-parental
populations for grain yield
The genetic gains per unit of time are given by the breeders’ equation,
which is Gain = (i·r·h)/I, where i is the selection intensity, r is the
selection accuracy, h is the square root of narrow-sense heritability, and
I is the time (in years) it takes to complete a selection cycle. In this
study, the gains in GY in different selection cycles were not consistent,
decreasing slightly from C0 to C2, while increasing significantly from C2
to C3, and from C3 to C4. As for analyses combining the two sites, the
gains in grain yield were 6.2 and 7.7% from C0 to C4 and from C1 to C4,
respectively. The combined realized genetic gains reached 0.158 ton
ha21 per cycle for C0–C4, (Figure 7A) and 0.225 ton ha21 per cycle
for RCGS C1–C4 (Figure 7B).
The lower GY observed in C1 compared to C0 is explained because
the best GY entries selected from C0 as parents of C1 were further Figure 6 Genome location of clusters of SNPs with changes in their
intermated and selected based on flowering, plant, and ear height, polymorphic status.

Volume 7 July 2017 | Genomic Selection in Multi-Parental Population | 2323

 Figure 7 Response to selection
for grain yield from the families
of (A) rapid cycling recombina-
tion genomic selection for cycles
C0, C1, C2, C3, and C4 and of (B)
rapid cycling recombination ge-
nomic selection for cycles C1,
C2, C3, and C4. Mean of the
checks (red), and mean of the
entries (blue).

Therefore, given that grain yield from C1 (8.40 ton ha21) to C4 drought environments (not optimal environments), and the RCGS in
(9.05 ton ha21) increased by 7.74%, the average genetic gain of this MPP targeted optimal environments.
0.225 ton ha21 per cycle (Table 2) is equivalent to 0.100 ton ha21 In this study, results obtained from MPPs in optimal environments
yr21 [i.e., (2 · 0.225)/4.5] under optimal conditions. reinforce the usefulness of GS-assisted recombination for achieving high
Masuka et al. (2017a) conducted a review of genetic gain studies that genetic gains in GY. Although only two cycles per year were completed in
used conventional pedigree selection on tropical hybrid maize germ- this study (Beyene et al. 2015, completed three cycles per year in bi-
plasm under optimal conditions in Sub-Saharan Africa, which gave parental populations), it is still time-efficient when compared to the
gains of 0.109 ton ha21 yr21. For tropical open-pollinated maize vari- 1.5 yr per selection cycle required for making testcrosses, phenotyping
testcrosses, and conducting selection and recombination in conven-
eties, realized genetic gains reached 0.109 ton ha21 yr21 in the early
tional pedigree breeding.
maturity group, and 0.079 ton ha21 yr21 in the intermediate-to-late
group (Masuka et al. 2017b). Therefore, the genetic gains from the RCGS
Trends in genetic diversity under rapid cycling
observed in the MPPs used in this study (0.100 ton ha21 yr21) are at recombination GS
the same or higher level than those observed in other studies under There are not many reports on the influence of RCGS on genetic
phenotypic selection but with a shorter breeding cycle. However, the variance in plant breeding. In a simulation study, Jannink et al.
0.070 ton ha21 yr21 achieved by Beyene et al. (2015) in bi-parental (2010) were the first to caution about the possible decline in genetic
populations is not comparable to the results of this study because the variance due to RCGS. Genetic gains in GS for stem rust in wheat were
genetic gains from RCGS in biparental populations targeted managed reported by Rutkoski et al. (2015); genetic gains in GS were compared

2324 | X. Zhang et al.

with gains in phenotypic selection and no differences were found. LITERATURE CITED
However, GS caused faster decline in genetic variance than phenotypic Ahfock, D., I. Wood, S. Stephen, C. R. Cavanagh, and B. E. Huang,
selection. Rutkoski et al. (2015) also found significant increases in in- 2014 Characterizing uncertainty in high-density maps from multipar-
breeding after one and two cycles of GS as compared with C0; this ental populations. Genetics 198: 117–128.
Bernardo, R., 2008 Molecular markers and selection for complex traits in
increase in inbreeding was significantly greater than the expected value
plants: learning from the last 20 years. Crop Sci. 48: 1649–1664.
under random genetic drift for all populations. Bernardo, R., 2016 Bandwagons I, too, have known. Theor. Appl. Genet.
The above results seem to be in partial agreement with the 129: 2323–2332.
findings of this study. The decrease in genetic diversity measured Bernardo, R., and J. Yu, 2007 Prospects for genome-wide selection for
by the Shannon Diversity Index and the expected heterozygosity quantitative traits in maize. Crop Sci. 47: 1082.
only occurred in RCGS in C3, whereas, in previous cycles, genetic Beyene, Y., K. Semagn, S. Mugo, A. Tarekegne, R. Babu et al., 2015 Genetic
diversity was very well maintained. These results may be due to the gains in grain yield through genomic selection in eight bi-parental maize
fact that C0 selection was initially based on GY, and then a second populations under drought stress. Crop Sci. 55: 154.
selection was conducted based on flowering, maturity, and other Cavanagh, C., M. Morell, I. Mackay, and W. Powell, 2008 From mutation
to MAGIC: resources for gene discovery, validation and delivery in crop
traits after intermating; this may be one of the reasons why the
plants. Curr. Opin. Plant Biol. 11: 215–221.
genetic diversity stayed at the same level as in C0, at least in the Crossa, J., P. Pérez-Rodríguez, J. Hickey, J. Burgueño, L. Ornella et al.,
initial recombination cycles. 2014 Genomic prediction in CIMMYT maize and wheat breeding
programs. Heredity 112: 48–60.
Changes in the frequency status of markers de Roos, A. P. W., B. J. Hayes, and M. E. Goddard, 2009 Reliability of
Results of this study show that only a few SNPs changed their poly- genomic predictions across multiple populations. Genetics 183: 1545–1553.
morphic status after three cycles of GS. This result indicates that marker Elshire, R. J., J. C. Glaubitz, Q. Sun, J. A. Poland, K. Kawamoto et al.,
interaction (epistasis) may play an important role in complex traits (such 2011 A robust, simple genotyping-by-sequencing (GBS) approach for
high diversity species. PLoS One 6: e19379.
as grain yield), in which phenotype is the result of the sum of the small
Glaubitz, J. C., T. M. Casstevens, F. Lu, J. Harriman, R. J. Elshire et al.,
effects of many genes. Only chromosome 5 showed a total of 19 clus- 2014 TASSEL-GBS: a high capacity genotyping-by-sequencing analysis
ters, with 16 of them having markers that became monomorphic (being pipeline. PLoS One 9: e90346.
polymorphic in their 18 parents and in the C0 training population); all Hoffstetter, A., A. Cabrera, M. Huang, and C. Sneller, 2016 Optimizing
the other chromosomes showed a small number of clusters with training population data and validation of genomic selection for eco-
markers that became monomorphic (from their original polymorphic nomic traits in soft winter wheat. G3 6: 2919–2928.
status) after cycles of RCGS. It is likely that some of those clusters of Huang, B. E. A., W. George, K. L. Forrest, A. Kilian, M. J. Hayden et al.,
SNP markers with changes in their polymorphic status are related to 2012 A multiparent advanced generation intercross population for ge-
netic analyses in wheat. Plant Biotechnol. J. 10: 826–839.
less complex traits that were selected during C0 and C1, such as flower-
Jannink, J.-L., A. J. Lorenz, and H. Iwata, 2010 Genomic selection in plant
ing time, and plant and ear height.
breeding from theory to practice. Brief. Funct. Genomics 9(2): 166–177.
Lehermeier, C., N. Kramer, E. Bauer, C. Bauland, C. Camisan et al.,
Conclusions 2014 Usefulness of multiparental populations of maize (Zea mays L.)
Results described in this study are the first report of RCGS in MPPs. A for genome-based prediction. Genetics 198: 3–16.
realized genetic gain of 2% for GY with two rapid cycles per year saves Massman, J. M., H.-J. G. Jung, and R. Bernardo, 2013 Genome-wide se-
time and produces efficient genetic gains overall. The decline in genetic lection vs. marker-assisted recurrent selection to improve grain yield and
gains from cycle C0 to C1 is because the aim of selecting parents for C1 stover-quality traits for cellulosic ethanol in maize. Crop Sci. 53: 58–66.
Masuka, B., G. Atlin, M. Olsen, C. Magorokosho, M. Labuschagne et al.,
was to maintain genetic diversity and the best selected families of cycle
2017a Gains in maize genetic improvement in Eastern and Southern
C0 were used as baseline for evaluating genetic gain among cycles. The Africa: I. CIMMYT hybrid breeding pipeline. Crop Sci. 57: 168.
realized gain achieved in this study was 0.100 ton ha21 yr21 when Masuka, B., C. Magorokosho, M. Olsen, G. Atlin, M. Bänziger et al.,
only GS cycles were considered (C1–C4). Another important finding is 2017b Gains in maize genetic improvement in Eastern and Southern
that genetic diversity was well controlled up to C3 and then declined. Africa: II. CIMMYT open-pollinated variety breeding pipeline. Crop Sci.
Although other traits were correlated with GY, they did not show any 57: 180.
important change after three cycles of RCGS for GY. In the end, 64.3% Meuwissen, T. H. E., B. J. Hayes, and M. E. Goddard, 2001 Prediction of
of the markers changed their allele frequencies but never became total genetic value using genome-wide dense marker maps. Genetics 157:
monomorphic, and 33.99% of the markers modified their allele fre- 1819–1829.
Rebai, A., and B. Goffinet, 1993 Power of tests for QTL detection using
quency by ,15%.
replicated progenies derived from a diallel cross. Theor. Appl. Genet. 86:
The target of this study was to perform three rapid cycles per year; 1014–1022.
however, we only achieved two rapid cycles per year due to delays in the Rutkoski, J., R. P. Singh, J. Huerta-Espino, B. Bhavani, J. Poland et al.,
DNA preparation and genotyping turnaround time. Therefore, further 2015 Genetic gain from phenotypic and genomic selection for quanti-
studies in maize or other crops are required to confirm the promising tative resistance to stem rust of wheat. Plant Genome 8: 2.
RCGS results obtained in MPP using tropical maize lines described in Swarts, K., H. Li, J. A. R. Navarro, D. An, M. C. Romay et al., 2014 Novel
this study, and examining if the length of selection cycle could be further methods to optimize genotypic imputation for low-coverage, next-
reduced by implementing RCGS. generation sequence data in crop plants. Plant Genome 7: 3.
Thepot, S., G. Restoux, I. Goldringer, F. Hospital, D. Gouache et al.,
2015 Efficiently tracking selection in a multiparental population: the
ACKNOWLEDGMENTS case of earliness in wheat. Genetics 199: 609–623.
The authors gratefully acknowledge the financial support from the VanRaden, P. M., 2007 Genomic measures of relationship and inbreeding.
MasAgro project funded by Mexico’s Secretariat of Agriculture, Live- Interbull Bull. 37: 33–36.
stock, Rural Development, Fisheries and Food (SAGARPA), and the VanRaden, P. M., 2008 Efficient methods to compute genomic predictions.
CGIAR Research Program (CRP) MAIZE. J. Dairy Sci. 91(11): 4414–4423.

Volume 7 July 2017 | Genomic Selection in Multi-Parental Population | 2325

Verhoeven, K. J. F., J.-L. Jannink, and L. M. McIntyre, 2006 Using mating Zhang, X., P. Pérez-Rodríguez, K. Semagn, Y. Beyene, R. Babu et al.,
designs to uncover QY+TL and the genetic architecture of complex traits. 2015 Genomic prediction in biparental tropical maize populations in
Heredity 96: 139–149. water-stressed and well-watered environments using low-density and
Vivek, B. S., G. K. Krishna, V. Vengadessan, R. Babu, P. H. Zaidi et al., GBS SNPs. Heredity 114: 291–299.
2017 Use of Genomic estimated breeding values results in rapid genetic Zhong, S., J. Dekkers, R. L. Fernando, and J. L. Jannink, 2009 Factors
gains for drought tolerance in maize. Plant Genome 10 (in press). affecting accuracy from genomic selection in populations derived from
Wu, Y., F. San Vicente, K. Huang, T. Dhliwayo, D. E. Costich et al., multiple inbred lines: a barley case study. Genetics 182: 355–364.
2016 Molecular characterization of CIMMYT maize inbred lines with
genotyping-by-sequencing SNPs. Theor. Appl. Genet. 129: 753–765. Communicating editor: D. J. de Koning

2326 | X. Zhang et al.