PLOS ONE
RESEARCH ARTICLE
Genomic selection for salinity tolerance in
japonica rice
Jérôme Bartholomé 1,2,3
ID *, Julien Frouin2,4, Laurent Brottier2,4, Tuong-Vi Cao2,4,
Arnaud Boisnard5, Nourollah Ahmadi2,4, Brigitte Courtois 2,4
ID
1 UMR AGAP Institut, CIRAD, Cali, Colombia, 2 UMR AGAP Institut, Institut Agro, Univ Montpellier, CIRAD,
INRAE, Montpellier, France, 3 Alliance Bioversity-CIAT, Recta Palmira Cali, Colombia, 4 CIRAD, UMR
AGAP Institut, Montpellier, France, 5 Centre Français du Riz, Arles, France
* jerome.bartholome@cirad.fr
a1111111111
a1111111111
a1111111111 Abstract
a1111111111
a1111111111 Improving plant performance in salinity-prone conditions is a significant challenge in breeding
programs. Genomic selection is currently integrated into many plant breeding programs as a
tool for increasing selection intensity and precision for complex traits and for reducing breeding
cycle length. A rice reference panel (RP) of 241 Oryza sativa L. japonica accessions geno-
typed with 20,255 SNPs grown in control and mild salinity stress conditions was evaluated at
OPEN ACCESS
the vegetative stage for eight morphological traits and ion mass fractions (Na and K). Weak to
Citation: Bartholomé J, Frouin J, Brottier L, Cao T-
strong genotype-by-condition interactions were found for the traits considered. Cross-valida-
V, Boisnard A, Ahmadi N, et al. (2023) Genomic
selection for salinity tolerance in japonica rice. tion showed that the predictive ability of genomic prediction methods ranged from 0.25 to 0.64
PLoS ONE 18(9): e0291833. https://doi.org/ for multi-environment models with morphological traits and from 0.05 to 0.40 for indices of
10.1371/journal.pone.0291833 stress response and ion mass fractions. The performances of a breeding population (BP) com-
Editor: Muhammad Abdul Rehman Rashid, prising 393 japonica accessions were predicted with models trained on the RP. For validation
Government College University Faisalabad, of the predictive performances of the models, a subset of 41 accessions was selected from the
PAKISTAN
BP and phenotyped under the same experimental conditions as the RP. The predictive abilities
Received: April 23, 2023 estimated on this subset ranged from 0.00 to 0.66 for the multi-environment models, depend-
Accepted: September 6, 2023 ing on the traits, and were strongly correlated with the predictive abilities on cross-validation in
Published: September 27, 2023 the RP in salt condition (r = 0.69). We show here that genomic selection is efficient for predict-
ing the salt stress tolerance of breeding lines. Genomic selection could improve the efficiency
Peer Review History: PLOS recognizes the
benefits of transparency in the peer review of rice breeding strategies for salinity-prone environments.
process; therefore, we enable the publication of
all of the content of peer review and author
responses alongside final, published articles. The
editorial history of this article is available here:
https://doi.org/10.1371/journal.pone.0291833 Introduction
Copyright: © 2023 Bartholomé et al. This is an Soil salinization is a major challenge worldwide, affecting about 20% of all irrigated land [1].
open access article distributed under the terms of
The predicted rise in sea level due to climate change will increase the percentage of salt-
the Creative Commons Attribution License, which
affected land, especially in coastal and delta areas [2]. Salinity is one of the most important fac-
permits unrestricted use, distribution, and
reproduction in any medium, provided the original tors reducing rice (Oryza sativa L.) crop productivity in river deltas, including some of the
author and source are credited. major production areas in Asia and Europe. This is a major concern for rice growers, as rice is
considered to be more salt-sensitive than other cereals, such as wheat or barley [3]. Indeed, salt
Data Availability Statement: The link for the data
availability statement is the following: https://doi. stress has a strong effect, decreasing growth or survival at the seedling stage in rice, even at
org/10.18167/DVN1/O1AYGP. moderate salinity levels, such as 3.4 dS�m−1 [4]. Salt stress also significantly impacts rice grain
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 1 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
Funding: This research was conducted under the yield, even at low salinity levels, with decreases in yield components, including tiller number,
framework of the FACCE-JPI project GreenRice spikelet number, and grain weight [5]. Sensitivity depends not only on the intensity of the
(Sustainable and environmental-friendly rice
stress, but also on its timing relative to plant development. Like other crops, rice is more sensi-
cultivation systems in Europe) and was funded by
tive during the seedling and reproductive stages [6]. Efforts have been made to unravel the
the French National Agency for Research (ANR-14-
JFAC-0005-01). The funders had no role in study mechanisms involved in salt tolerance in rice at the physiological, molecular, and genetic levels
design, data collection and analysis, decision to [7,8]. Three mechanisms have been implicated in plant salt-stress tolerance: ion exclusion
publish, or preparation of the manuscript. (Na+ or Cl−), tissue tolerance, and osmotic tolerance [6,9]. Phenotyping is challenging due to
Competing interests: The authors declare that they the complex interactions between environmental and genetic factors, and accurate protocols
have no competing interests. are required to quantify plant responses to salt stress [6].
Since the start of the Green Revolution, rice breeders and geneticists have been screening
Abbreviations: RP, reference panel; BP, breeding
population; SNP, single nucleotide polymorphism; the diversity of O. sativa, with the aim of identifying tolerant genotypes [10,11]. Screening pro-
BP41, a subset of 41 lines of the BP; CRTL, control tocols have been developed to evaluate lines subjected to different stress levels at the seedling
condition; GBLUP, Genomic best linear unbiased stage [12]. In the 1970s, for example, the International Rice Research Institute in the Philip-
prediction; GBS, genotyping by sequencing; G×E, pines evaluated about 100,000 accessions for salt tolerance by visual damage scoring. Less than
genotype by environment interactions; LA, leaf
20% of the lines tested were found to be tolerant, suggesting that salt tolerance is not a major
area; LL, leaf length; PA, predictive ability; QTL,
quantitative trait locus; R_S, ratio of root-to-shoot characteristic in rice [13,14]. Most of the tolerant accessions belonged to the O. sativa subspe-
dry weights; RKHS, reproducing kernel Hilbert cies indica (e.g., Nona Bokra, Pokkali, Hasawi, Getu, Cheriviruppu, Ddamodar, Solla and
space; RL, root length; ROOT, root dry weight; RP, Ketumbar) and many originated from mangrove areas, but some tolerant lines were also
reference panel; SALT, salt condition; SHOOT, found in the japonica subspecies (e.g., Honduras, Slava, and Gigante Vercelli, [15,16]). These
shoot dry weight; SLA, specific leaf area; SNP,
tolerant landraces or ecotypes were subsequently used as donors for salinity tolerance genes in
single nucleotide polymorphism; TIL, number of
breeding programs [13]. With the advent of molecular markers, tolerant landraces were used
tillers.
in various types of crosses, to map quantitative trait loci (QTLs) and to shed light on the
genetic control of salinity tolerance [17–19]. Several hundred QTLs and dozens of genes have
been identified, highlighting the complex genetic control of this trait [12,13]. These QTLs and
genes are spread throughout the rice genome, but with a higher concentration on chromosome
1. The major QTL SalTol, located on chromosome 1, was used for marker-assisted selection
[20–23]. Most efforts have focused on pre-breeding activities based on QTL introgression, and
a number of salt-tolerant varieties, such as IR64-Saltol, ASS996-Saltol and BRRI dhan 47, have
emerged from these approaches [24,25]. However, most of these varieties belong to the indica
subspecies, which grows in tropical and subtropical regions. In the framework of the European
Neurice project, Saltol was introgressed in Spanish, Italian and French temperate japonica
varieties through marker-assisted selection(http://www.neurice.eu).
Despite major advances in the marker-assisted selection of a major QTL/gene, breeding for
salt tolerance is rendered more difficult by the complex genetic architecture of the trait, as hun-
dreds of genes associated with different mechanisms are involved [13,26]. Favorable alleles at
several QTLs/genes are therefore required to occur together to confer a significant level of toler-
ance in field conditions. There is a need to combine QTL/genes controlling salt tolerance at
both the vegetative and reproductive stages. In addition, interactions with environmental condi-
tions, such as the timing and intensity of salt stress, play a major role, making it harder to iden-
tify the best-performing lines. Indeed, tolerant lines must also perform well under more
favorable conditions, and must be adapted to the various environments encountered in farmers’
fields [27]. In this context, genomic selection is a potential tool for accelerating genetic improve-
ment for salt tolerance [28]. Genomic selection requires the initial calibration of a prediction
model on a training population that has been both genotyped and phenotyped. This prediction
model is then applied to candidates for selection, based purely on their genotypes [29]. This
approach enables breeders to optimize their breeding strategy by selecting early in the breeding
cycle for complex and expensive-to-evaluate traits, thereby decreasing the length of the breeding
cycle [30]. In rice, genomic prediction has been successfully evaluated for different traits and
different types of populations [31–33]. Most of these studies focused on predicting
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 2 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
performances in normal conditions, but a few have assessed the accuracy of prediction for per-
formance under abiotic stress conditions, such as water deficit in particular [34,35]. Genomic
prediction models integrating genotype-by-environment interactions have recently been devel-
oped [36–39]. These models tend to increase prediction accuracy, by incorporating multi-envi-
ronment data and modeling marker-by-environment interactions. They are particularly useful
for breeding programs targeting stress-prone environments, as they make it possible to combine
performances in the presence and absence of stress, to increase accuracy [31].
In this study, we used material from European rice breeding programs to evaluate the poten-
tial of genomic prediction for improving salinity tolerance in japonica rice adapted to temperate
regions. Due to environmental constraints (low temperatures and long days) almost all of the
rice varieties grown in Europe belong to the japonica subspecies [40]. A panel of European
accessions was recently evaluated for tolerance to salt stress at the seedling stage, revealing con-
siderable variability for various traits (growth parameters and sodium (Na+) and potassium
(K+) mass fractions) despite the absence of the favorable allele at the Saltol locus [41]. These
findings indicate that the use of minor genes and selection for this quantitative variation by
genomic selection could help to increase the salt tolerance of temperate japonica breeding lines.
However, for most breeding programs in temperate regions, only one growing season per year
is possible, and the evaluation of salinity tolerance in field conditions is always difficult because
of the high level of interannual variability. Genomic prediction models are, therefore, useful for
predicting the salinity tolerance of untested genotypes, to make it easier for breeders to take
selection decisions as early as possible. The objectives of this study were: i) to assess the accuracy
of single and multi-environment genomic prediction models for predicting the performance of
accessions from European rice breeding programs for traits related to salt tolerance via cross-
validation and ii) to validate these models on an independent subset of breeding lines.
Materials and methods
Plant material
We used two different populations (S1 Table). The first population was a reference panel (RP)
composed of 241 japonica accessions. The RP was previously characterized by Frouin, et al.
[41]. These accessions were mostly varieties from temperate regions, with European accessions
largely represented (Italy (46.7%), France (12.5%), Spain (12.5%), Portugal (7.1%)), alongside
varieties from the United States (10.4%), and more than 15 other countries. The second popu-
lation was a breeding population (BP) of 393 breeding lines, with 73.8% of the lines derived
from the joint breeding program of the Centre Français du Riz (CFR, Arles, France) and the
Centre de cooperation international en recherche agronomique pour le développement (CIRAD,
Montpellier, France). This French material included 289 current advanced breeding lines
derived from 98 crosses involving 114 parents, 33 of which were included in the RP. The num-
ber of lines per cross ranged from 1 to 20, with seven crosses overrepresented (121 individu-
als). The other lines in the BP were lines from the working collections of European breeders.
Genotypic characterization of the populations
The two populations were genotyped with the same genotyping-by-sequencing method [42].
The DNA was extracted with the a modified method using hexadecyltrimethylammonium
bromide [43], as described by Frouin et al. [41]. The genome was digested with the restriction
enzyme ApeKI for library preparation. The libraries were sequenced with a Genome Analyzer
II (Illumina, San Diego, California, USA). The Nipponbare reference genome Os-Nipponbare-
Reference-IRGSP-1.0, [44] was used for sequence alignment. Single-nucleotide polymorphism
(SNP) calling was performed with the Tassel GBS pipeline with the default parameters [45].
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 3 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
Markers with a call rate below 75%, a heterozygosity rate above 10% and a minor allele fre-
quency below 5.0% were discarded. The remaining heterozygotes were converted to missing
data. The missing data were then imputed with Beagle v4.0, using the default parameters [46].
The imputed file was split into two sets: the 241 accessions of the RP and the 393 accessions of
the BP. Markers with a minor allele frequency below 5.0% were discarded in both populations.
This procedure resulted in the identification of 20,255 informative SNPs common to the two
populations. Markers in complete linkage disequilibrium were further filtered out: for clusters
of markers, only one marker, that with the lowest rate of missing data before imputation and
the highest minor allele frequency, was selected to represent the cluster. This filter resulted in
16,993 non-redundant markers, which were used for further analysis. The data can be down-
loaded from CIRAD Dataverse: https://doi.org/10.18167/DVN1/O1AYGP.
The genetic structure of the 241 accessions of the RP and 393 lines of the BP was assessed
by a discriminant analysis of principal components implemented in the R package adegenet
[47,48]. The functions find.clusters and dapc were successively used to assign individuals to
genetic groups. The number of clusters was set to three, corresponding to the tropical and tem-
perate japonica subgroups and admixed accessions. The percentage of the variance used to
select the number of axes in the principal component analysis was set at 90%. We also used
DarWin v6 software [49] to calculate a simple matching index to assess the dissimilarity
between individuals. We also used DarWin v6 software [49] to calculate a simple matching
index to assess the dissimilarity between individuals. This index was then used to construct an
unweighted neighbor-joining tree. The assignments to groups derived from adegenet were
projected onto the neighbor-joining tree.
Assessment of phenotypic performance under controlled conditions
Reference panel (RP). The accessions of the RP were phenotyped for salinity tolerance
under hydroponic conditions, as previously described by Frouin et al. [41]. Briefly, the experi-
mental design was a split-plot with three replicates staggered in time. In each replicate, the
plants were distributed in 12 tanks (6 control and 6 salt tanks). Two resistant controls (Nona
Bokra and Pokkali) and three susceptible controls (IR29, Aychade and Giano) were replicated
in each tank. Stress was applied for two weeks, beginning two weeks after sowing. The salt con-
centration was 50 mM NaCl (3 g/l), corresponding to an electrical conductivity of 6.5 dS�m-1.
After 28 days sowing, we measured the following growth-related traits in control and salt con-
ditions: the number of tillers (TIL), the lengths of the longest leaf (LL) and the longest root
(RL), the length (LGTH) and width (WDTH) of the last fully developed leaf of the main tiller,
the dry matter weight for shoots (SHOOT), roots (ROOT) and the last fully developed leaf of
the main tiller (LEAF). The Na+ and K+ mass fractions of shoot tissues, expressed as percent of
dry matter, were measured on plants grown in salt conditions, by atomic emission spectros-
copy (ICP-AES), at the UR59 laboratory at CIRAD Montpellier (ISO9001). We also calculated
the following variables: leaf area (LA), calculated as LGTH x WIDTH x 0.75, specific leaf area
(SLA), calculated as LA divided by LEAF, the root-to-shoot ratio (R/S), calculated as ROOT
divided by SHOOT, and the Na/K ratio, calculated as Na+ mass fraction over K+ mass fraction.
Analysis of variance was performed on the morphological trait data, with a mixed model
including salinity/control conditions and genotype as fixed effects and replicate and tank as
random effects. The significance of genotype, conditions and genotype x conditions interac-
tion effects were assessed. The least square mean values of the genotypes were calculated with
SAS software (Cary, NC, USA).
From the least square mean values for morphological traits, stress response indices were
computed as iTRAIT = (salt−control)×100/control.
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 4 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
Subset of the breeding population. The same methodology was used to phenotype a sub-
set of 41 lines selected from the BP (BP41) in a separate experiment. The design was a split plot
with two conditions (control and salt stress) and three replicates. The resistant and susceptible
controls varieties were as described above. The same statistical model as for the RP was used
for the variance analysis except that, in this case, because of the smaller size of the experiment,
the tank and replicate effects were combined. The process used to select the 41 lines is
explained below in the “Evaluation of predictive ability” section.
Statistical models for genomic prediction
The genomic BLUP (GBLUP) and the reproducing kernel Hilbert space (RKHS) models were
used to predict breeding values with molecular markers. GBLUP is one of the most popular
and robust methods for genomic prediction [50,51]. For GBLUP, we calculated the kernel
matrix as follows K = XX’/p, X being the centered genotype matrix. X is of dimension n×p,
where n is the number of genotypes and p the number of markers. For the RKHS model, we
used a Gaussian kernel Kðxi; xjÞ ¼ expð hk xi  xj k2Þ to calculate the kernel matrix between
the marker genotype vectors xi and xj, where (i,j)2{1,. . .,N}2. We estimated the bandwidth
parameter h as described by Pérez-Elizalde et al. [52] and with the associated R function
margh.fun. This method is based on estimation of the mode of the joint posterior distribution
of h and a form parameter φ. The shape and scale parameters of the gamma prior distribution
for h were 3.0 and 1.5, respectively.
The extensions of the GBLUP and RKHS models for multivariate analysis were used to pre-
dict the genomic estimated breeding values from data for the two sets of conditions. This
approach is referred to hereafter as “multi-environment prediction”. For both the extended
GBLUP and RKHS models, the effects of markers are divided into two components: the main
effect and the environment-specific effects [38,53]. Following the notation of Cuevas et al.
[38], the model can be expressed as:
2 3 2 3 2 3 2 32 3 2 3
y
1 1m1 X1 X1 . . . 0 . . . 0 β1 ε1
6 7 6 7 6 7 6 76 7 6 7
6 . 7 6 . 7 6
6 .. 7 6 . 7 .. 7 6
6 7 .. . . .
6
6 7 . . . .. . . .. 76 7 6
6 7 6 . 7 . . 6 ..7 7 .. 76 7
6 76 . 7 6 . 7
6 7 6 7 6 7 6 76 7 6 7
6 7
6 yj 7 ¼ 6 7
6 1mj 7þ
6 7
6 Xj 7β0 þ
6 76 7
6 0 . . . Xj . . . 0 76 βj 7þ
6 7
6 εj 7
6 7 6 7 6 7 6 76 7 6 7
6 7 6
..
6 . 7 6
6 7 . . 7 .. .
4 . 6 7 6 . 7
4 5 . . . .. .
. . . ..
76 7 6 . 7
6 ..
5 4 . 5 . 6 . 7 7 6 .
4 54 . 5 4 . 75
ym 1mm Xm 0 . . . 0 . . . Xm βm εm
Where yj is the response vector in the jth environment, μj is the intercept in the jth environ-
ment, Xj, is the centered matrix of marker in the jth environment, β0 is the vector of marker
effects across all environments, βj is the vector of marker effects for environment j and εj is the
random error for the jth environment. By using the following notation:
2 3 2 3 2 3 2 32 3 2 3
y
1 1m1 X1 X1 . . . 0 . . . 0 β1 ε1
6 7 6 7 6 7 6 76 7 6 7
6 7 6 7
6 .. 6
7 .. 7 .
6 7 .. . . .
6 . 6 7 . 6
7 6 . 7 6 . 7 . . . .. . . .. 76 . 7 6
6 7 . 6 .. 76 7
6 . 76 . 7 6 . 77
6 7 6 7 6 7 6 76 7 6 7
y ¼ 6 7 6 7
6 yj 7; m ¼ 6 7
6 1mj 7; u0 ¼ 6 Xj 7β 0 . . . X
0; u 6
E ¼ j . . . 0 76 7
6 76 βj 7; ε ¼ 6 7
6 εj 7
6 7 6 7 6 7 6 76 7 6 7
6 7 6 7
6 ..
6 7 6
... 7 6 7 ..6 7 .. . . .
4 5 4 . 5 4 . 5 . . . . . 76 7
6 7 ..
6
7 ..
7
. . . . 6
4 . 54 . 6 7
5 4 . 5
ym 1mm Xm 0 . . . 0 . . . Xm βm εm
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 5 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
the model can be written as follow:
y ¼ mþ uo þ uE þ ε
In this mixed model, the random effect uo follows a multivariate normal distribution with a
mean of zero and a variance–covariance matrix K0s
2
u with K0 ¼ X0X
0
0=p for the GBLUP and
o
K0ðx0i; x0jÞ ¼ expð h0k x0i  x0j k2Þ for RKHS. The random effect uE follows a multivariate
normal distribution with a mean of zero and a variance–covariance matrix KE with
2
s2
u K
3
. . .
1 1 . . 0 . 0
6 7
6 .. . . .
. . . .. . . .. 7
6 7
6 . 7
6 7
KE ¼
6 0 . . . s2
u K . . . 0 7
6 j j 7. The kernel matrix for each environment is esti-
6 7
6 7
6 .. .
. . .. . .
. . . .. 7
4 . 5
0 . . . 0 . . . s2
u K1 m
mate as follow: Kj ¼ XjX
0
j=p for the GBLUP and Kjðxji; xjjÞ ¼ expð hjk xji  xjj k2Þ for RKHS
Analyses were performed in the R 3.6.1 environment [54]. The GBLUP and RKHS models
were fitted in the R package BGLR 1.0.8 [55]. Inferences were based on 3,000 of the 35,000 iter-
ations for the Gibbs sampler. The first 5,000 samples were discarded (burn-in) and we then
kept one sample out of ten to avoid autocorrelation (thinning). Convergence of the Markov
Chain-Monte Carlo algorithm was assessed for all parameters of the models, with Gelman-
Rubin tests [56] and the R package coda 0.19–1 [57].
Evaluation of predictive ability
Cross-validation within the reference panel. We estimated the predictive ability (PA) of
the models described above using a cross-validation strategy within the RP: we randomly
selected 80% of the panel to form the training set, with the remaining 20% used as the valida-
tion set. For multi-environment models, the genotypes composing the training set were associ-
ated with phenotypic information for the two sets of conditions, whereas no phenotypic
information was available for those composing the validation set. This cross-validation
approach is usually referred to as CV1 in the literature [58]. The random partitioning of the
RP was repeated 100 times, and the PA for each partition was calculated as the Pearson coeffi-
cient of correlation between the genomic estimated breeding values and the corresponding
phenotypes in the validation set. For each combination of model (single or multi-environ-
ment), statistical method (GBLUP, RKHS) and trait, the same partitions were used to calculate
predictive ability. The resulting estimates of predictive ability were averaged, and the associ-
ated standard error was calculated. We analyzed the effect of the different factors (trait, condi-
tions, prediction method, etc.) on PA, by performing analyses of variance. To avoid potential
bias due to the distribution of the coefficient of correlation (r2[−1; 1]), we transformed it
using Fisher Z transformation according to the following equation:
Z ¼ 0:5fln½1þ r�  ln½1  r�g. The analyses of variance were done on the Z statistics.
Validation in the breeding population (BP). We evaluated PA in the breeding popula-
tion in two steps. We first used the 241 accessions of RP to train the model and then predicted
the phenotype of each of the 393 lines in the BP. We then selected a set of 41 lines from the BP
for actual phenotyping under the same conditions as the RP. The number of lines selected was
chosen to allow for the phenotyping of three replicates in hydroponic conditions. This set of
41 lines (BP41) was chosen to be representative of the variability of the predicted phenotypic
values of the most important salt tolerance traits observed for the 393 lines of the BP—Na, K
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 6 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
and Na/K—and for iSHOOT and iROOT, predicted by both the GBLUP and RKHS methods.
We captured the variability present in the BP, for each trait, by selecting lines ranked in the
lowest 10%, lines ranked in the top 10% and the lines with average performances in the follow-
ing proportion: 20%, 20% and 60%. PA was calculated as the Pearson coefficient of correlation
between the predicted phenotypes of the 41 lines obtained with models trained on the RP and
the corresponding actual phenotypes.
Results
Characterization of the reference panel and the breeding population
In total, 20,255 informative SNPs spread throughout the genome were common to the RP and
the BP (S1 Fig). The mean distance between SNPs was 18.3 kb, and the largest gap between
markers was 1.3 Mb, on chromosome 11. Forty-two gaps of more than 500 kb were observed
throughout the genome, with chromosomes 3 and 5 presenting the largest numbers of such
gaps. SNPs in complete linkage disequilibrium (r2 = 1) were removed. This resulted in 16,993
markers with a distribution very similar to the initial dataset (S1 Fig). The distribution of minor
allele frequencies across non-redundant markers was similar for the RP and the BP (S2 Fig).
Analyses of the genetic structures of both the RP and BP highlighted the well-known bipolar
structure of European rice accessions, with temperate and tropical japonica subgroups (S3 Fig).
As expected, given the nature of the genetic material, the level of admixture between these two
subgroups was high, reaching 37% in both the RP and the BP. The BP was highly related to the
RP, as highlighted by the small genetic distance between the two populations (Fig 1). This relat-
edness between the two populations is a key parameter for genomic prediction.
Fig 1. Unweighted neighbor-joining tree representing the dissimilarities between individuals composing the reference
panel (black) and the breeding population (blue and green for the subset used for validation).
https://doi.org/10.1371/journal.pone.0291833.g001
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 7 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
The 241 accessions of the RP had been evaluated under two hydroponic conditions: control
and salt. Weak-to-moderate genotype x conditions interactions (G×C) were observed, depend-
ing on the trait (Fig 2). The weakest G×C interaction was that for leaf length (LL), with a rank
correlation (Kendall) between the two conditions of ρ = 0.79. Conversely, specific leaf area
(SLA) presented the strongest G×C interaction, with ρ = 0.33. The 10% of accessions with the
Fig 2. Phenotypic performance of the accessions of the reference panel in the two sets of hydroponic conditions (CRTL: Control; SALT: Salt).
All traits measured in both conditions are presented: Number of tillers (TIL), leaf length (LL), leaf area (LA), specific leaf area (SLA), root length
(RL), root dry weight (ROOT), shoot dry weight (SHOOT) and the ratio of root-to-shoot dry weights (R_S). The accessions in green are the 10% of
accessions with the lowest Na/K ratios in salt conditions. Spearman’s rank correlation coefficients (ρ) between control and salt conditions are
indicated at the top of each panel.
https://doi.org/10.1371/journal.pone.0291833.g002
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 8 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
lowest Na/K ratios did not differ significantly (p-value > 0.05) from the rest of the RP for any
combination of trait/conditions other than root/shoot ratio (R_S) in salt conditions, for which
the ratio was higher for the rest of the panel. The correlation between morphological traits ran-
ged from -0.45 to 0.90 in control conditions and from -0.37 to 0.90 in salt-stress conditions
(S2 Table). ROOT was strongly correlated with SHOOT (0.90) and LL was significantly corre-
lated with ROOT, SHOOT and LA, in both sets of conditions. However, the correlations
between traits tended to be weaker in salt conditions than in control conditions. In salt condi-
tions, the ion mass fractions presented few significant correlations with morphological traits:
NA/K was correlated with TIL (-0.20), SHOOT (-0.19) and R_S (0.22, S2 Table).
The phenotypic variability observed in the RP was partly related to population structure.
Indeed, Na/K was significantly higher for the admixed subgroup than for the tropical and tem-
perate subgroups (S4 Fig). A difference was also found for RL, for which the admixed sub-
group had lower values, suggesting a higher susceptibility to the effects of salinity on root
development (S4 Fig).
Predictive ability of genomic prediction within the reference panel
Comparison between single and multi-environment models. Cross-validation in the RP
gave predictive ability (PA) values for the multi-environment model ranging from 0.25 to 0.64
in control conditions and from 0.38 to 0.63 in salt conditions, for the eight morphological
traits analyzed (Fig 3). Similar PAs were obtained for the eight traits in the single environment
Fig 3. Estimates of predictive ability for performances in the two sets of conditions (CTRL and SALT) for the
reference panel. Performances were predicted with multi-environment (black) or single-environment (gray) models.
Two different prediction methods were used: GBLUP and RKHS. The traits considered were: Number of tillers (TIL),
leaf length (LL), leaf area (LA), specific leaf area (SLA), root length (RL), root dry weight (ROOT), shoot dry weight
(SHOOT) and the ratio of root-to-shoot dry weights (R_S). The bars represent the average predictive ability over 100
replicates, and the error bars represent the standard error of the mean.
https://doi.org/10.1371/journal.pone.0291833.g003
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 9 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
model, with values ranging from 0.22 to 0.63 in control conditions and from 0.36 to 0.62 in
salt conditions. For both the single-environment and multi-environment models, the lowest
PA was that for SLA and the highest was that for LL in control conditions. PA was higher in
control conditions than in salt conditions for six of the eight traits (TIL, LL, LA, SHOOT,
ROOT, and R_S) (Fig 3). The largest positive difference in accuracy (0.19) between control
and salt conditions was for LA when the single environment model was used. For SLA and RL,
PA was higher under salt conditions than under control conditions, with differences up to 0.24
and 0.17 (single-environment model, RKHS), respectively. All the factors considered in the
analysis of variance (trait, conditions, prediction method and prediction model) had a signifi-
cant effect on PA, with the largest effect for trait, the other factors being of a much lesser
importance (S3 Table). Indeed, differences in PA between single and multi-environment mod-
els were close to zero, except for SLA and RL in control conditions for which accuracy was
15% and 10% higher, respectively, for the multi-environment model than for the single-envi-
ronment model. Interestingly, these were the two traits with the best PA in salt conditions. No
clear relationship was found between the strength of G×C interactions for traits and the per-
formance of the multi-environment model relative to that of the single-environment model.
Stress response indices and ion mass fractions. The stress response indices combined
values from both control and salt conditions, whereas mass fractions were measured only
under salt conditions. Only the single environment model could, therefore, be used for mass
fractions. The PA values for the indices were lower than those of other traits in each set of con-
ditions, with values ranging from -0.05 to 0.35 (Fig 4A). On average, over the two prediction
methods, the iROOT, iSHOOT and iLL indices had the lowest PAs (lower than 0.10). Con-
versely, the PAs for iRL and iSLA were greater than 0.30. The PA for Na and K mass fractions
and for Na/K ranged from 0.20 to 0.40 (Fig 4A). The K mass fraction was slightly better pre-
dicted than Na mass fraction, with PAs of 0.34 (GBLUP) and 0.40 (RKHS), versus 0.28
(GBLUP) and 0.29 (RKHS), respectively. Both trait and prediction method had a significant
effect on PA. The effect of trait was the most significant, with prediction method having only a
marginal effect (S4 Table). A negative correlation was found between PA and the strength of
G×C interactions for the trait: -0.69 (p-value = 0.059) and -0.58 (p-value = 0.134) for GBLUP
and RKHS, respectively (Fig 4B).
Validation of predictive ability in the breeding population
The performances of the 393 genotypes of the BP were predicted with the two prediction mod-
els (single and multi-environment) built on the RP. As expected, the genomic estimated per-
formances were shrunk toward the mean value of the RP, and this effect was more
pronounced for RKHS than for GBLUP. Depending on the trait, the coefficients of correlation
(ρ) between the predicted performances estimated with RKHS and GBLUP ranged from 0.88
(SLA) to 0.98 (LL) for the multi-environment model and from 0.76 (SLA) to 0.99 (LL) for the
single-environment model (S5 Table). For indices and ion mass fractions, the coefficients of
correlation ranged from 0.29 (iROOT) to 0.97 (iLA). The traits presenting the lowest correla-
tion between prediction methods were those with the lowest PAs in cross-validation in the RP:
iROOT, iSHOOT, and iLL. For Na and K, and for Na/K, the coefficients of correlation
between prediction methods exceeded 0.90.
For validation of the predicted performances, a subset of 41 lines from the BP (BP41) was
selected for phenotyping. This selection had little effect on the extent of variability of the pre-
dicted traits relative to the entire BP (S5 and S6 Figs), but it decreased the neutral genetic vari-
ability, as shown by the distribution of BP41 on the neighbor-joining tree (Fig 1). The
repeatability of the actual phenotypic data obtained for BP41 through evaluation under control
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 10 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
Fig 4. (A) Predictive abilities for stress response indices (iTrait) and ion mass fraction (Na and K) in the reference
panel. Two different prediction methods are presented: GBLUP in orange and RKHS in gray. A Tukey boxplot is used
to represent the data. (B) Scatterplot of estimates of predictive ability for indices and the level of genotype-by-
environment interactions estimated by calculating the Spearman correlation coefficient for the same genotype in
control and salt conditions.
https://doi.org/10.1371/journal.pone.0291833.g004
and salt conditions ranged from 0.70 to 0.93 for the measured traits and from 0.44 to 0.55 for
the indices (Table 1). Under salt conditions, the selected lines displayed phenotypic variation
for ion mass fraction (Na and K) within the expected range relative to the susceptible and tol-
erant controls (Fig 5). G×C interactions of similar strength (ρ) to those in the RP were found
(Table 1). The weakest G×C interaction was that for LA, with a rank correlation ρ between the
two conditions of 0.83, and the strongest G×C interaction was that for SLA (ρ = 0.25).
The phenotypic data for BP41 were used to estimate the PA of the models trained on the RP
for the various traits. For the multi-environment model, PA ranged from 0.00 to 0.66 and
from 0.09 to 0.54 in control and salt conditions, respectively (Table 1). As for cross-validation,
similar results were obtained with the single-environment model, with values ranging from
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 11 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
Table 1. Summary of statistics for the 41 lines of the breeding population (BP41) for the various traits measured under control and salt conditions.
Prediction TIL LL LA SLA RL ROOT SHOOT R_S Na* K* Na/
method K*
H2 (control/salt) - 0.70 / 0.89 / 0.88/ 0.48/ 0.88 / 0.82 / 0.89 / 0.87 / 0.82 0.85 0.82
0.73 0.93 0.87 0.07 0.69 0.86 0.89 0.76
ρ - 0.38 0.66 0.83 0.25 0.67 0.31 0.54 0.80 - - -
PA with multi-environment model RKHS 0.02 / 0.38 / 0.36 / 0.50 / 0.63 / 0.66 / 0.4 0.43 / 0.48 / - - -
(control/salt) 0.14 0.56 0.50 0.14 0.45 0.43 0.31
GBLUP 0.00 / 0.43 / 0.36 / 0.57 / 0.64 / 0.65 / 0.42 / 0.47 / - - -
0.10 0.54 0.45 0.09 0.36 0.29 0.34 0.41
PA with single-environment model RKHS 0.03 / 0.37 / 0.37 / 0.45 / 0.61 / 0.66 / 0.46 / 0.48 / - - -
(control/salt) 0.14 0.59 0.53 0.17 0.42 0.45 0.47 0.28
GBLUP 0.02 / 0.39 / 0.37 / 0.56 / 0.64 / 0.67 / 0.46 / 0.41 / - - -
0.04 0.61 0.49 0.12 0.32 0.36 0.40 0.40
PA for the indices and ion content RKHS -0.05 0.21 0.11 0.35 0.19 0.33 0.34 0.02 0.31 0.26 0.26
GBLUP 0.01 0.06 0.13 0.32 0.31 0.26 0.33 -0.07 0.32 0.35 0.25
Repeatability in the two conditions (H2), the rank correlation between conditions (ρ),predictive abilities (PA) for single- and multi-environment models and for the
indices are provided. The traits presented are: Number of tillers (TIL), leaf length (LL), leaf area (LA), specific leaf area (SLA), root length (RL), root dry weight (ROOT),
shoot dry weight (SHOOT), the ratio of root-to-shoot dry weights (R_S), the ion mass fractions of Na and K and their ratio (Na/K). * The ion mass fractions were
measured only in salt conditions.
https://doi.org/10.1371/journal.pone.0291833.t001
0.02 to 0.67 in control conditions and from 0.04 to 0.61 in salt conditions. For stress-response
indices and ion mass fractions, PA ranged from -0.07 to 0.35. The PAs for the Na (0.31 and
0.32, with RKHS and GBLUP, respectively) and K (0.26 and 0.35) mass fractions (the main
traits used to select BP41) were intermediate, lying between those for the other traits. The dif-
ferences between RKHS and GBLUP were slightly larger than those observed for cross-valida-
tion. Depending on the trait and the model, gains in PA of up to 0.10 were observed for RKHS
(SHOOT in salt conditions) or for GBLUP (SLA in control conditions). Interestingly, the PAs
estimated by cross-validation and those obtained with the subset of the BP were not correlated
for control conditions, whereas there was a non-significant trend towards a positive correla-
tion between these PAs in salt conditions (Fig 6, S6 Table).
Discussion
Performance of japonica accessions under salt stress
In this study, we evaluated the salt tolerance of accessions and advanced lines from European
breeding programs at the seedling stage under hydroponic conditions, with a no-salt control
and a salt-stress treatment of 50 mM NaCl (6.5 dS�m-1). Hydroponic experiments are very use-
ful for evaluations of the sensitivity of accessions to a given level of salt stress [6,59]. This
approach has been used extensively to screen tolerant material in breeding programs [11] as
salt levels are highly variable in field experiments, with micro-environmental and seasonal var-
iations that can bias the evaluation of accessions. However, field-based evaluations and screen-
ing in controlled conditions are jointly used in breeding programs as stress tolerance at the
seedling stage may not fully correlate with field performance. In our experiments, a moderate
stress level, corresponding to a degree of salinity commonly observed in the Camargue region
of France was used [60]. Considerable variability was observed in the phenotypic response to
salt stress in both the RP and BP41, consistent with the findings of previous studies on japonica
accessions [15,16,61]. In their work on 176 temperate japonica accessions, Batayeva et al. [61]
reported that only a few accessions (Nep Ngau, Bai Mang Ai Zhong, and Shinchiku Iku 97)
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 12 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
Fig 5. Distribution of the ion mass fractions for Na and K in the subset of the breeding population (BP41). The
two tolerant controls (Nona Bokra and Pokkali) are represented in green and the three susceptible controls (IR29,
Giano and Aychade) are shown in pink. The lines from BP41 are in gray. The error bars represent the confidence
interval of the mean (95%) which was calculated based on the data for the three replicates for each line.
https://doi.org/10.1371/journal.pone.0291833.g005
were as tolerant as the control variety, the others being moderately tolerant or susceptible. In
this study, we found that three lines had performances similar to that of the tolerant control,
with low Na_K values under salt conditions (Fulgente, Escarlate and RX110_01).
Accuracy of genomic predictions for salt tolerance
Genomic prediction has been studied in detail in rice in recent years, with more than 50 stud-
ies published to date [33]. However, none of these studies focused on predicting genotype per-
formances under salt stress. Depending on the trait, the type of population, the validation
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 13 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
Fig 6. Comparison of the predictive ablities for eight morphological traits estimated by cross-validation on the
reference panel and the 41 selected lines from the breeding population used as a validation set. Control conditions
are shown in black, and salt conditions in gray. Spearman’s rank correlation coefficients (ρ) are indicated.
https://doi.org/10.1371/journal.pone.0291833.g006
method and the prediction method, the accuracies reported in these studies varied substan-
tially, from close to zero to more than 0.90. Here, prediction accuracy ranged from 0.05 to 0.63
for the various traits when the cross-validation method was applied to the RP. No clear differ-
ences were found between the two prediction methods used (GBLUP and RKHS). These
results are consistent with those reported in previous studies [34,51,62,63]. We found that the
differences between single- and multi-environment models were also small, although multi-
environment models outperformed single-environment models for most of the traits and
tended to be more accurate for predictions in salt conditions. The small differences between
single- and multi-environment models were expected, given the type of cross-validation
scheme used, with the prediction of untested genotypes only. With this type of cross-validation
scheme, multi-environment models and single-environment models tend to perform similarly
[34,64,65]. Multi-environment models have been shown to perform better when the predicted
genotypes are evaluated in at least one environment (mimicking sparse testing evaluation). In
this case, the gain in prediction accuracy of multi-environment models over single-environ-
ment models may be between 30% and 50%, as shown by Ben Hassen et al. [34] with pheno-
typic data for two sets of conditions (continuous flooding and alternate wetting and drying).
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 14 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
Cuevas et al. [38] reported similar results for a wheat dataset including four environments,
with gains of up to 60%, depending on the traits considered and the validation method.
Stress response indices had lower predictive abilities than regular traits. For the indices, pre-
dictive ability was negatively correlated with the strength of G×C interactions. This result
reflects the difficulty in predicting and selecting for relative performance between a stress envi-
ronment (in our case salt stress) and normal conditions. The positive relationship between
predictive ability and heritability is well known and has been described in studies on genomic
prediction [66,67], but the impact of genotype-by-environment interactions on predictive abil-
ity tends to be more difficult to characterize [68,69]. Furthermore, for indices, the errors asso-
ciated with measurements in the two sets of conditions are propagated in the index, tending to
decrease heritability. Ion mass fractions (Na and K) and their ratio (Na/K) had low-to-moder-
ate predictive abilities (0.2 to 0.4) with similar values in both populations (RP and BP41).
These predictive abilities are in the expected range given the complex genetic architecture
underlying these two traits [41]. No other study on rice has assessed the performance of geno-
mic prediction in the context of salinity tolerance. We were therefore unable to make direct
comparisons for this trait. However, the predictive abilities obtained in this study for Na and K
could potentially be improved with prediction methods modeling marker effects differently
from GBLUP and RKHS. Methods such as Bayesian LASSSO, BayesB and Random Forest
have been shown to perform better when only a few regions of the genome have moderate
effect on the target traits [51,70,71]. Similarly, the integration of GWAS results into the geno-
mic prediction model might also increase prediction accuracy [35,72,73]. In this study, no
major QTL was found in the RP [41], but this approach may be of interest when major QTLs,
such as Saltol are segregating in the breeding population.
Importance of validation on selection candidates
Predictive ability or accuracy is routinely used to evaluate the efficiency of genomic prediction
models [74]. In most genomic selection studies in plants, accuracy is measured as the correla-
tion (r) between observed and predicted phenotypic performances [75]. The most common
approach for estimating accuracy is cross-validation (subset validation), because of its ease of
use. In this approach, the dataset is split in two (the calibration set and the validation set), mak-
ing it possible to estimate accuracies in a given population while keeping parameters such as
marker density, population structure or allele frequencies constant. However, cross-validation
tends to overestimate accuracy relative to other validation approaches, such as inter-set valida-
tion or progeny validation [63,76,77]. Most of the studies performed in rice have used cross-
validation to obtain estimates of prediction accuracy [33]. Here, we used both cross-validation
and inter-set validation (validation in the breeding population). The two approaches gave sim-
ilar estimates of predictive ability. The accuracies obtained by cross-validation and with BP41
were well-correlated, but only in salt conditions. This finding reflects the method used to select
lines for inclusion in BP41: ion mass fractions measured under salt conditions. The use of sim-
ilar phenotyping conditions for model training and validation is therefore important, to obtain
a more precise idea of the level of accuracy that can be expected, as GxE interactions generally
decrease accuracy [69,78]. The relatedness between the RP and the BP, with a similar structure
in the two populations, may also explain this result. Indeed, the genetic distance between the
training set and the validation set has been shown to be one of the major factors affecting accu-
racy [79,80]. BP41 constitutes only a small subset of the entire BP, but most of the parental
accessions and closely related lines, were present in the RP used to train the model. Ben Hassen
et al. [63] reported similar results in a study using a diversity panel to predicted advance mate-
rial from the breeding program.
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 15 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
Implications for breeding for salt tolerance in rice
Genotypes tolerant to salt stress are clearly less common among japonica accessions than
among indica accessions [15,16]. However, it is important for temperate rice breeding pro-
grams to characterize their material better for salinity tolerance because the sources of toler-
ance identified in indica accessions are difficult to introgress due to the complex genetic
architecture of the trait, as revealed by linkage mapping [81] and association studies
[16,41,82]. Sterility problems have also been reported for intersubspecific indica x japonica
crosses [83,84]. We show here that genomic selection can be used to predict the salt tolerance
of material from European rice breeding programs (japonica) under mild constraints. The
major challenge would be efficiently combining evaluations under normal and salt conditions.
As discussed above, capturing GxE interactions (or GxC interactions) with multi-environment
models can increase the accuracy of predictions. Genomic selection for salt tolerance can be
implemented just after line fixation (usually F6). In breeding programs involving rapid genera-
tion advances, such selection can take place in the third year of the breeding strategy [33]. The
use of genomic selection at this stage would make it possible to select lines on the basis of their
predicted performance in both normal and salt conditions. Multi-environment genomic pre-
diction could be used to combine information from normal and salt conditions (either in the
greenhouse or in the field) for prediction in a larger set of candidates from the cohort of the
next cycle. The selection intensity for salt tolerance would be increased early in the breeding
scheme (e.g., stage 1 yield trial). As stress trials are difficult to manage, a targeted set of lines
could be evaluated in stress conditions, but with a higher degree of replication, to update the
model.
Supporting information
S1 Fig. Distribution in the rice genome of the informative markers for the complete set of
20,255 SNPs (upper panel) and the non-redundant set of 16,993 SNPs (lower panel).
(PDF)
S2 Fig. Distribution of minor allele frequency (MAF) for the 16,993 non-redundant SNPs
in the two populations: The reference panel and the breeding population.
(PDF)
S3 Fig. Unweighted neighbor-joining tree and the associated genetic structure estimated
for K = 2 in the upper panel and K = 3 in the lower panel. Temperate japonica is shown in
red, tropical japonica in blue and admixed accessions are shown in purple.
(PDF)
S4 Fig. Boxplot for the stress response indices (iTrait) and the K and Na mass fractions
and their ratio in the reference panel. The different subpopulations (admixed, temperate,
tropical) were defined with molecular markers (see materials and methods). Different letters
for a given trait indicate a significant difference between group means (Tukey’s HSD test,
p< 0.05).
(PDF)
S5 Fig. Boxplot of genomic estimated breeding value (GEBV) for the eight morphological
traits in the breeding population of 393 lines. The 41 lines selected for the validation experi-
ment are represented in black and the rest of the population is shown in gray. Two prediction
methods (GBLUP and RKHS) and two models (single- and multi-environment) were com-
pared.
(PDF)
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 16 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
S6 Fig. Distribution of genomic estimated breeding value (GEBV) for Na and K mass frac-
tion and their ratio (Na/K) in the breeding population of 393 lines. The 41 breeding lines
selected for the validation experiment are represented in green and the rest of the population
is shown in gray. Two different prediction methods were used: GBLUP and RKHS.
(PDF)
S7 Fig. Assessment of salinity tolerance under hydroponic conditions. The image at the top
represents half the 12 tanks for one replicate for the reference panel. The image at the bottom
represents the three replicates for the selected genotypes of the breeding population. The con-
trol tanks are shown on the left and those for salt conditions are shown on the right.
(PDF)
S1 Table. List of accessions in the two populations: The reference panel and the breeding
population.
(XLSX)
S2 Table. Correlations between morphological traits in control conditions (upper table)
and between ion mass fractions in salt conditions (lower table). The Spearman rank correla-
tion coefficients are displayed in the lower part of the matrices and the associated p-value are
shown in the upper part.
(PDF)
S3 Table. Analysis of variance of predictive abilities in the reference panel for perfor-
mances in both sets of conditions (CTRL and SALT). Two prediction methods were com-
pared (GBLUP and RKHS), for eight traits and two models (single- and multi-environment).
(PDF)
S4 Table. Analysis of variance of predictive abilities in the reference panel for indices and
ion mass fractions (referred to as Trait in the table). Two prediction methods were com-
pared (GBLUP and RKHS).
(PDF)
S5 Table. Spearman’s rank correlation coefficient for the relationship between the pre-
dicted performances estimated with RKHS and GBLUP for the entire breeding population,
with single- and multi-environment models.
(PDF)
S6 Table. Relationship between predictive abilities estimated by cross-validation on the
reference panel and those estimated with the subset (41 lines) of the breeding population.
Two models (single- and multi-environment) and two methods (GBLUP and RKHS) were
evaluated.
(PDF)
Acknowledgments
This work has been realized with the support of MESO@LR-Platform at the University of
Montpellier. The authors thank US49 from CIRAD for conducting the mass fraction analyses.
Author Contributions
Conceptualization: Nourollah Ahmadi, Brigitte Courtois.
Data curation: Jérôme Bartholomé, Julien Frouin, Laurent Brottier, Tuong-Vi Cao, Arnaud
Boisnard, Brigitte Courtois.
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 17 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
Formal analysis: Jérôme Bartholomé, Julien Frouin, Brigitte Courtois.
Funding acquisition: Nourollah Ahmadi, Brigitte Courtois.
Investigation: Jérôme Bartholomé.
Methodology: Jérôme Bartholomé, Brigitte Courtois.
Project administration: Nourollah Ahmadi, Brigitte Courtois.
Supervision: Brigitte Courtois.
Visualization: Jérôme Bartholomé.
Writing – original draft: Jérôme Bartholomé.
Writing – review & editing: Julien Frouin, Laurent Brottier, Tuong-Vi Cao, Arnaud Boisnard,
Nourollah Ahmadi, Brigitte Courtois.
References
1. Metternicht GI, Zinck JA. Remote sensing of soil salinity: potentials and constraints. Remote Sensing of
Environment. 2003; 85(1):1–20.
2. IPCC. Climate Change 2013: the Physical Science Basis. Contribution of Working Group I to the Fifth
Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge: Cambridge Uni-
versity Press; 2013. Report No.: 9780521880091 0521880092 9780521705967 0521705967.
3. Maas EV, Grattan SR. Crop Yields as Affected by Salinity. Agricultural Drainage 1999. p. 55–108.
4. Zeng L, Shannon MC. Salinity Effects on Seedling Growth and Yield Components of Rice. Crop Sci-
ence. 2000; 40(4):996–1003.
5. Grattan S, Zeng L, Shannon M, Roberts S. Rice is more sensitive to salinity than previously thought.
California agriculture. 2002; 56(6):189–98.
6. Negrão S, Schmöckel SM, Tester M. Evaluating physiological responses of plants to salinity stress.
Annals of Botany. 2016; 119(1):1–11. https://doi.org/10.1093/aob/mcw191 PMID: 27707746
7. Munns R, James RA, Läuchli A. Approaches to increasing the salt tolerance of wheat and other cereals.
Journal of Experimental Botany. 2006; 57(5):1025–43. https://doi.org/10.1093/jxb/erj100 PMID:
16510517
8. Ismail AM, Horie T. Genomics, Physiology, and Molecular Breeding Approaches for Improving Salt Tol-
erance. Annual Review of Plant Biology. 2017; 68(1):405–34. https://doi.org/10.1146/annurev-arplant-
042916-040936 PMID: 28226230
9. Horie T, Karahara I, Katsuhara M. Salinity tolerance mechanisms in glycophytes: An overview with the
central focus on rice plants. Rice. 2012; 5(1):11. https://doi.org/10.1186/1939-8433-5-11 PMID:
27234237
10. Akbar M, Shakoor A, Sajjad MS. Prospects of Breeding for Salt Tolerance in Rice. In: Muhammed A,
Aksel R, von Borstel RC, editors. Genetic Diversity in Plants. Boston, MA: Springer US; 1977. p. 291–
9.
11. Gregorio G, Senadhira D, Mendoza R. Screening rice for salinity tolerance. Los Banos: International
Rice Research Institute; 1997 01/01.
12. Singh RK, Kota S, Flowers TJ. Salt tolerance in rice: seedling and reproductive stage QTL mapping
come of age. Theor Appl Genet. 2021; 134(11):3495–533. https://doi.org/10.1007/s00122-021-03890-
3 PMID: 34287681
13. Negrão S, Courtois B, Ahmadi N, Abreu I, Saibo N, Oliveira MM. Recent Updates on Salinity Stress in
Rice: From Physiological to Molecular Responses. Critical Reviews in Plant Sciences. 2011; 30
(4):329–77.
14. Platten JD, Egdane JA, Ismail AM. Salinity tolerance, Na+ exclusion and allele mining of HKT1;5 in
Oryza sativa and O. glaberrima: many sources, many genes, one mechanism? BMC Plant Biology.
2013; 13(1):32. https://doi.org/10.1186/1471-2229-13-32 PMID: 23445750
15. Lee K-S, Choi W-Y, Ko J-C, Kim T-S, Gregorio GB. Salinity tolerance of japonica and indica rice (Oryza
sativa L.) at the seedling stage. Planta. 2003; 216(6):1043–6. https://doi.org/10.1007/s00425-002-
0958-3 PMID: 12687373
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 18 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
16. Ahmadi N, Negrão S, Katsantonis D, Frouin J, Ploux J, Letourmy P, et al. Targeted association analysis
identified japonica rice varieties achieving Na+/K+ homeostasis without the allelic make-up of the salt
tolerant indica variety Nona Bokra. Theor Appl Genet. 2011; 123(6):881–95. https://doi.org/10.1007/
s00122-011-1634-4 PMID: 21713536
17. Flowers TJ, Koyama ML, Flowers SA, Sudhakar C, Singh KP, Yeo AR. QTL: their place in engineering
tolerance of rice to salinity. Journal of Experimental Botany. 2000; 51(342):99–106. PMID: 10938800
18. Singh R, Gregorio GB, Jain R. QTL mapping for salinity tolerance in rice. Physiology and Molecular Biol-
ogy of Plants. 2007; 13(2):87.
19. Jena KK, Mackill DJ. Molecular markers and their use in marker-assisted selection in rice. Crop Sci-
ence. 2008; 48(4):1266–76.
20. Thomson MJ, de Ocampo M, Egdane J, Rahman MA, Sajise AG, Adorada DL, et al. Characterizing the
Saltol Quantitative Trait Locus for Salinity Tolerance in Rice. Rice. 2010; 3(2):148–60.
21. Bimpong IK, Manneh B, Sock M, Diaw F, Amoah NKA, Ismail AM, et al. Improving salt tolerance of low-
land rice cultivar ‘Rassi’ through marker-aided backcross breeding in West Africa. Plant Science. 2016;
242:288–99. https://doi.org/10.1016/j.plantsci.2015.09.020 PMID: 26566846
22. Punyawaew K, Suriya-arunroj D, Siangliw M, Thida M, Lanceras-Siangliw J, Fukai S, et al. Thai jasmine
rice cultivar KDML105 carrying Saltol QTL exhibiting salinity tolerance at seedling stage. Molecular
Breeding. 2016; 36(11):150.
23. Linh LH, Linh TH, Xuan TD, Ham LH, Ismail AM, Khanh TD. Molecular Breeding to Improve Salt Toler-
ance of Rice (Oryza sativa L.) in the Red River Delta of Vietnam. International Journal of Plant Geno-
mics. 2012; 2012:949038. https://doi.org/10.1155/2012/949038 PMID: 23326259
24. Huyen LTN, Cuc LM, Ismail AM, Ham LH. Introgression the Salinity Tolerance QTLs &lt;i&gt;Saltol&lt;/
i&gt; into AS996, the Elite Rice Variety of Vietnam. American Journal of Plant Sciences. 2012;Vol.
03No.07:4.
25. Salam MA, Rahman M, Bhuiyan MAR, Uddin K, Sarker MRA, Yasmeen R, et al. BRRI dhan 47: a salt-
tolerant variety for the boro season. International Rice Research Notes. 2007; 32:42–3.
26. Reddy INBL Kim B-K, Yoon I-S Kim K-H, Kwon T-R. Salt Tolerance in Rice: Focus on Mechanisms and
Approaches. Rice Science. 2017; 24(3):123–44.
27. Niang A, Becker M, Ewert F, Dieng I, Gaiser T, Tanaka A, et al. Variability and determinants of yields in
rice production systems of West Africa. Field Crops Research. 2017; 207:1–12.
28. Morton MJL, Awlia M, Al-Tamimi N, Saade S, Pailles Y, Negrão S, et al. Salt stress under the scalpel–
dissecting the genetics of salt tolerance. The Plant Journal. 2019; 97(1):148–63. https://doi.org/10.
1111/tpj.14189 PMID: 30548719
29. Meuwissen THE, Hayes BJ, Goddard ME. Prediction of Total Genetic Value Using Genome-Wide
Dense Marker Maps. Genetics. 2001; 157(4):1819–29. https://doi.org/10.1093/genetics/157.4.1819
PMID: 11290733
30. Cobb JN, Juma RU, Biswas PS, Arbelaez JD, Rutkoski J, Atlin G, et al. Enhancing the rate of genetic
gain in public-sector plant breeding programs: lessons from the breeder’s equation. Theor Appl Genet.
2019. https://doi.org/10.1007/s00122-019-03317-0 PMID: 30824972
31. Ahmadi N, Bartholomé J, Tuong-Vi C, Grenier C. Genomic selection in rice: empirical results and impli-
cations for breeding. Wallingford: CABI; 2020. p. 243–58.
32. Spindel J, Iwata H. Genomic Selection in Rice Breeding. In: Sasaki T, Ashikari M, editors. Rice Geno-
mics, Genetics and Breeding. Singapore: Springer Singapore; 2018. p. 473–96.
33. Bartholomé J, Prakash PT, Cobb JN. Genomic Prediction: Progress and Perspectives for Rice Improve-
ment. In: Ahmadi N, Bartholomé J, editors. Complex Trait Prediction: Methods and Protocols. New
York, NY: Springer US; 2022. p. 569–617.
34. Ben Hassen M, Bartholome J, Vale G, Cao TV, Ahmadi N. Genomic Prediction Accounting for Geno-
type by Environment Interaction Offers an Effective Framework for Breeding Simultaneously for Adap-
tation to an Abiotic Stress and Performance Under Normal Cropping Conditions in Rice. G3-Genes
Genomes Genetics. 2018; 8(7):2319–32. https://doi.org/10.1534/g3.118.200098 PMID: 29743189
35. Bhandari A, Bartholomé J, Cao-Hamadoun T-V, Kumari N, Frouin J, Kumar A, et al. Selection of trait-
specific markers and multi-environment models improve genomic predictive ability in rice. PLOS ONE.
2019; 14(5):e0208871. https://doi.org/10.1371/journal.pone.0208871 PMID: 31059529
36. Jarquı́n D, Crossa J, Lacaze X, Du Cheyron P, Daucourt J, Lorgeou J, et al. A reaction norm model for
genomic selection using high-dimensional genomic and environmental data. TAG Theoretical and
Applied Genetics Theoretische Und Angewandte Genetik. 2014; 127(3):595–607. https://doi.org/10.
1007/s00122-013-2243-1 PMID: 24337101
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 19 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
37. Montesinos-López A, Montesinos-López OA, Crossa J, Burgueño J, Eskridge KM, Falconi-Castillo E,
et al. Genomic Bayesian Prediction Model for Count Data with Genotype × Environment Interaction.
G3: Genes|Genomes|Genetics. 2016; 6(5):1165–77.
38. Cuevas J, Crossa J, Soberanis V, Pérez-Elizalde S, Pérez-Rodrı́guez P, Campos Gdl, et al. Genomic
Prediction of Genotype × Environment Interaction Kernel Regression Models. The Plant Genome.
2016; 9(3).
39. Cuevas J, Crossa J, Montesinos-López OA, Burgueño J, Pérez-Rodrı́guez P, de los Campos G. Bayes-
ian Genomic Prediction with Genotype × Environment Interaction Kernel Models. G3: Genes|Genomes|
Genetics. 2017; 7(1):41–53.
40. Courtois B, Frouin J, Greco R, Bruschi G, Droc G, Hamelin C, et al. Genetic Diversity and Population
Structure in a European Collection of Rice. Crop Science. 2012; 52(4):1663–75.
41. Frouin J, Languillaume A, Mas J, Mieulet D, Boisnard A, Labeyrie A, et al. Tolerance to mild salinity
stress in japonica rice: A genome-wide association mapping study highlights calcium signaling and
metabolism genes. PLOS ONE. 2018; 13(1):e0190964. https://doi.org/10.1371/journal.pone.0190964
PMID: 29342194
42. Elshire RJ, Glaubitz JC, Sun Q, Poland JA, Kawamoto K, Buckler ES, et al. A Robust, Simple Genotyp-
ing-by-Sequencing (GBS) Approach for High Diversity Species. PLOS ONE. 2011; 6(5):e19379. https://
doi.org/10.1371/journal.pone.0019379 PMID: 21573248
43. Gawel NJ, Jarret RL. A modified CTAB DNA extraction procedure forMusa andIpomoea. Plant Mol Biol
Rep. 1991; 9(3):262–6.
44. Kawahara Y, de la Bastide M, Hamilton JP, Kanamori H, McCombie WR, Ouyang S, et al. Improvement
of the Oryza sativa Nipponbare reference genome using next generation sequence and optical map
data. Rice. 2013; 6(1):4. https://doi.org/10.1186/1939-8433-6-4 PMID: 24280374
45. Glaubitz JC, Casstevens TM, Lu F, Harriman J, Elshire RJ, Sun Q, et al. TASSEL-GBS: A High Capac-
ity Genotyping by Sequencing Analysis Pipeline. PLOS ONE. 2014; 9(2):e90346. https://doi.org/10.
1371/journal.pone.0090346 PMID: 24587335
46. Browning SR, Browning BL. Rapid and Accurate Haplotype Phasing and Missing-Data Inference for
Whole-Genome Association Studies By Use of Localized Haplotype Clustering. The American Journal
of Human Genetics. 2007; 81(5):1084–97. https://doi.org/10.1086/521987 PMID: 17924348
47. Jombart T, Devillard S, Balloux F. Discriminant analysis of principal components: a new method for the
analysis of genetically structured populations. BMC Genetics. 2010; 11(1):94. https://doi.org/10.1186/
1471-2156-11-94 PMID: 20950446
48. Jombart T. adegenet: a R package for the multivariate analysis of genetic markers. Bioinformatics.
2008; 24(11):1403–5. https://doi.org/10.1093/bioinformatics/btn129 PMID: 18397895
49. Perrier X, Jacquemoud-Collet J. DARwin software. http://darwinciradfr/. 2006.
50. Van Raden PM. Efficient Methods to Compute Genomic Predictions. Journal of Dairy Science. 2008; 91
(11):4414–23. https://doi.org/10.3168/jds.2007-0980 PMID: 18946147
51. Azodi CB, Bolger E, McCarren A, Roantree M, de los Campos G, Shiu S-H. Benchmarking Parametric
and Machine Learning Models for Genomic Prediction of Complex Traits. G3: Genes|Genomes|Genet-
ics. 2019; 9(11):3691–702.
52. Pérez-Elizalde S, Cuevas J, Pérez-Rodrı́guez P, Crossa J. Selection of the Bandwidth Parameter in a
Bayesian Kernel Regression Model for Genomic-Enabled Prediction. Journal of Agricultural, Biological,
and Environmental Statistics. 2015; 20(4):512–32.
53. Lopez-Cruz M, Crossa J, Bonnett D, Dreisigacker S, Poland J, Jannink J-L, et al. Increased Prediction
Accuracy in Wheat Breeding Trials Using a Marker × Environment Interaction Genomic Selection
Model. G3: Genes|Genomes|Genetics. 2015; 5(4):569–82.
54. R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical
Computing, Vienna, Austria: URL http://www.R-project.org/; 2018.
55. Pérez P, de los Campos G. Genome-Wide Regression and Prediction with the BGLR Statistical Pack-
age. Genetics. 2014; 198(2):483–95. https://doi.org/10.1534/genetics.114.164442 PMID: 25009151
56. Gelman A, Rubin DB. Inference from iterative simulation using multiple sequences. Statistical science.
1992:457–72.
57. Plummer M, Best N, Cowles K, Vines K. CODA: convergence diagnosis and output analysis for MCMC.
R news. 2006; 6:7–11.
58. Burgueño J, de los Campos G, Weigel K, Crossa J. Genomic Prediction of Breeding Values when
Modeling Genotype × Environment Interaction using Pedigree and Dense Molecular Markers. Crop Sci-
ence. 2012; 52(2):707–19.
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 20 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
59. Bado S, Forster BP, Ghanim AMA, Jankowicz-Cieslak J, Berthold G, Luxiang L. Protocol for Screening
for Salt Tolerance in Rice. Protocols for Pre-Field Screening of Mutants for Salt Tolerance in Rice,
Wheat and Barley. Cham: Springer International Publishing; 2016. p. 21–31.
60. Audebert A, Mendez Del Villar P. Characterization of rice crop systems and rice sector organisation in
Camargue, France. Montpellier: CIRAD; 2005.
61. Batayeva D, Labaco B, Ye C, Li X, Usenbekov B, Rysbekova A, et al. Genome-wide association study
of seedling stage salinity tolerance in temperate japonica rice germplasm. BMC Genetics. 2018; 19
(1):2. https://doi.org/10.1186/s12863-017-0590-7 PMID: 29298667
62. Bustos-Korts D, Malosetti M, Chapman S, Biddulph B, van Eeuwijk F. Improvement of Predictive Ability
by Uniform Coverage of the Target Genetic Space. G3-Genes Genomes Genetics. 2016; 6(11):3733–
47. https://doi.org/10.1534/g3.116.035410 PMID: 27672112
63. Ben Hassen M, Cao TV, Bartholome J, Orasen G, Colombi C, Rakotomalala J, et al. Rice diversity
panel provides accurate genomic predictions for complex traits in the progenies of biparental crosses
involving members of the panel. Theor Appl Genet. 2018; 131(2):417–35. https://doi.org/10.1007/
s00122-017-3011-4 PMID: 29138904
64. Monteverde E, Rosas JE, Blanco P, de Vida FP, Bonnecarrere V, Quero G, et al. Multienvironment
Models Increase Prediction Accuracy of Complex Traits in Advanced Breeding Lines of Rice. Crop Sci-
ence. 2018; 58(4):1519–30.
65. Jarquı́n D, Lemes da Silva C, Gaynor RC, Poland J, Fritz A, Howard R, et al. Increasing Genomic-
Enabled Prediction Accuracy by Modeling Genotype × Environment Interactions in Kansas Wheat. The
Plant Genome. 2017.
66. Elsen J-M. An analytical framework to derive the expected precision of genomic selection. Genetics
Selection Evolution. 2017; 49(1):95. https://doi.org/10.1186/s12711-017-0366-6 PMID: 29281960
67. Resende MFR, Muñoz P, Resende MDV, Garrick DJ, Fernando RL, Davis JM, et al. Accuracy of Geno-
mic Selection Methods in a Standard Data Set of Loblolly Pine (Pinus taeda L.). Genetics. 2012; 190
(4):1503–10. https://doi.org/10.1534/genetics.111.137026 PMID: 22271763
68. Malosetti M, Bustos-Korts D, Boer MP, van Eeuwijk FA. Predicting Responses in Multiple Environ-
ments: Issues in Relation to Genotype × Environment Interactions. Crop Science. 2016; 56(5):2210–
22.
69. Crossa J, Pérez-Rodrı́guez P, Cuevas J, Montesinos-López O, Jarquı́n D, de los Campos G, et al.
Genomic Selection in Plant Breeding: Methods, Models, and Perspectives. Trends in Plant Science.
2017; 22(11):961–75. https://doi.org/10.1016/j.tplants.2017.08.011 PMID: 28965742
70. Spindel J, Begum H, Akdemir D, Virk P, Collard B, Redoña E, et al. 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 Elite, Tropical Rice
Breeding Lines. PLoS Genet. 2015; 11(2):e1004982. https://doi.org/10.1371/journal.pgen.1004982
PMID: 25689273
71. Heslot N, Yang H-P, Sorrells ME, Jannink J-L. Genomic Selection in Plant Breeding: A Comparison of
Models. Crop Sci. 2012; 52(1):146–60.
72. Spindel JE, Begum H, Akdemir D, Collard B, Redoña E, Jannink J-L, et al. Genome-wide prediction
models that incorporate de novo GWAS are a powerful new tool for tropical rice improvement. Heredity.
2016; 116(4):395–408. https://doi.org/10.1038/hdy.2015.113 PMID: 26860200
73. Ahmadi N, cao T-V, Frouin J, Norton GJ, Price AH. Genomic Prediction of Arsenic Tolerance and Grain
Yield in Rice: Contribution of Trait-Specific Markers and Multi-Environment Models. Rice Science.
2021; 28(3):268–78.
74. Elsen J-M. Genomic Prediction of Complex TraitsComplex traits, Principles, Overview of Factors Affect-
ing the Reliability of Genomic PredictionGenomic predictions, and Algebra of the Reliability. In: Ahmadi
N, Bartholomé J, editors. Complex Trait Prediction: Methods and Protocols. New York, NY: Springer
US; 2022. p. 45–76.
75. Lorenz AJ, Chao S, Asoro FG, Heffner EL, Hayashi T, Iwata H, et al. Genomic Selection in Plant Breed-
ing. Knowledge and Prospects.2011 2011. 77 p.
76. Michel S, Ametz C, Gungor H, Epure D, Grausgruber H, Löschenberger F, et al. Genomic selection
across multiple breeding cycles in applied bread wheat breeding. Theor Appl Genet. 2016; 129
(6):1179–89. https://doi.org/10.1007/s00122-016-2694-2 PMID: 27067826
77. Sallam AH, Endelman JB, Jannink J-L, Smith KP. Assessing Genomic Selection Prediction Accuracy in
a Dynamic Barley Breeding Population. The Plant Genome. 2015; 8(1).
78. Crossa J, Montesinos-López OA, Pérez-Rodrı́guez P, Costa-Neto G, Fritsche-Neto R, Ortiz R, et al.
Genome and Environment Based Prediction Models and Methods of Complex Traits Incorporating
Genotype × Environment Interaction. Methods Mol Biol. 2022; 2467:245–83.
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 21 / 22
PLOS ONE Genomic selection for salinity tolerance in japonica rice
79. Scutari M, Mackay I, Balding D. Using Genetic Distance to Infer the Accuracy of Genomic Prediction.
PLOS Genetics. 2016; 12(9):e1006288. https://doi.org/10.1371/journal.pgen.1006288 PMID:
27589268
80. Rio S, Charcosset A, Mary-Huard T, Moreau L, Rincent R. Building a Calibration Set for Genomic Pre-
dictionGenomic predictions (GP), Characteristics to Be Considered, and OptimizationOptimizationsAp-
proaches. In: Ahmadi N, Bartholomé J, editors. Complex Trait Prediction: Methods and Protocols. New
York, NY: Springer US; 2022. p. 77–112.
81. Islam MS, Ontoy J, Subudhi PK. Meta-Analysis of Quantitative Trait Loci Associated with Seedling-
Stage Salt Tolerance in Rice (Oryza sativa L.). Plants. 2019; 8(2):33. https://doi.org/10.3390/
plants8020033 PMID: 30699967
82. Chen C, Norton GJ, Price AH. Genome-Wide Association Mapping for Salt Tolerance of Rice Seedlings
Grown in Hydroponic and Soil Systems Using the Bengal and Assam Aus Panel. Frontiers in Plant Sci-
ence. 2020;11.
83. Liu KD, Zhou ZQ, Xu CG, Zhang Q, Saghai Maroof MA. An analysis of hybrid sterility in rice using a dial-
lel cross of 21 parents involving indica, japonica and wide compatibility varieties. Euphytica. 1996; 90
(3):275–80.
84. Oka H-I. Genic analysis for the sterility of hybrids between distantly related varieties of cultivated rice.
Journ of Genetics. 1957; 55(3):397–409.
PLOS ONE | https://doi.org/10.1371/journal.pone.0291833 September 27, 2023 22 / 22