Gene-set association and epistatic analyses reveal complex gene interaction networks affecting flowering time in a worldwide barley collection

Using gene-set association test and epistasis analysis, this research achieved higher statistical power with potentially high accuracy, and detected significant genes and gene networks that influence flowering time in barley.

Abstract

Single-marker genome-wide association studies (GWAS) have successfully detected associations between single nucleotide polymorphisms (SNPs) and agronomic traits such as flowering time and grain yield in barley. However, the analysis of individual SNPs can only account for a small proportion of genetic variation, and can only provide limited knowledge on gene network interactions. Gene-based GWAS approaches provide enormous opportunity both to combine genetic information and to examine interactions among genetic variants. Here, we revisited a previously published phenotypic and genotypic data set of 895 barley varieties grown in two years at four different field locations in Australia. We employed statistical models to examine gene–phenotype associations, as well as two-way epistasis analyses to increase the capability to find novel genes that have significant roles in controlling flowering time in barley. Genetic associations were tested between flowering time and corresponding genotypes of 174 putative flowering time-related genes. Gene–phenotype association analysis detected 113 genes associated with flowering time in barley, demonstrating the unprecedented power of gene-based analysis. Subsequent two-way epistasis analysis revealed 19 pairs of gene×gene interactions involved in controlling flowering time. Our study demonstrates that gene-based association approaches can provide higher capacity for future crop improvement to increase crop performance and adaptation to different environments.

Introduction

Barley (Hordeum vulgare L.) is one the most important cereal crops in the world and is cultivated both in highly productive agricultural regions and in marginal environments prone to adverse conditions (Baum et al., 2007). As a particularly resilient crop compared with other cereals such as wheat and rice, barley has the ability to adapt to biotic and abiotic stresses, holding a great deal of potential to increase production in marginal areas to sustain food security (Tester and Langridge, 2011). It is vital that barley flowers within a particular time window in a given environment to maximize yield, while minimizing exposure to frost, heat, and drought stress during the growing season (Maurer et al., 2015). It is also known that genes controlling phenology including flowering time (FT) overlap with many grain yield-related genes (Hill et al., 2019a). Sharma et al. (2018) identified a total of 96 quantitative trait loci (QTLs) mapped for grain yield in a nested association mapping population, the majority of which also co-localized with known genes controlling FT.

Harnessing the power of genomic tools to manipulate FT for barley improvement is of considerable importance to meet the food and feed demands of the future. Understanding the genetic basis of FT including the interactions between different FT genes has the potential to considerably enhance genetic improvement and future barley breeding. Insights gained from model plants such as Arabidopsis thaliana made it possible to explore the function of gene orthologues and related pathways in barley, but not all genes and gene networks discovered in A. thaliana are conserved across the plant kingdom. For example, monocot-specific genes and gene networks, including species-specific flowering gene networks in rice, have been reported (Xue et al., 2008; Matsubara et al., 2011).

Rapid advancements in genome sequencing technologies including reduced representation sequencing approaches, combined with high-throughput genotyping and the availability of a high-quality reference genome, now allow for an unprecedented view into complex genetic architectures in barley (Waugh et al., 2009; Huang et al., 2011; Mayer et al., 2012; Mascher et al., 2017; Sharma et al., 2018; Hill et al., 2019a). Genome-wide association studies (GWAS) have emerged as powerful tools for identifying genetic variants associated with crop plant phenotypes (Pasam et al., 2012; Yano et al., 2016; Fang et al., 2017). Commonly used single-marker GWAS approaches test each single nucleotide polymorphism (SNP) individually for the association with a trait, which has delivered considerable insight into the genetic control of traits (Yang et al., 2014). However, only the most significant SNPs in the genome are taken into account with the single-marker approach, thus can often explain only a small proportion of the genetic variation. In fact, single SNP variants explained <10% of phenotypic variation for the majority of complex phenotypes (Manolio et al., 2009). Moreover, single SNP analyses consider only the effect of individual SNP and often examine additive models only, while most quantitative traits are polygenic and thus also determined by gene×gene interactions (epistasis).

Epistasis is known to play a crucial role in regulation of many complex traits in plants, animals, and humans (Doust et al., 2014; Phillips, 2008). Different theoretical frameworks and statistical methodologies for epistasis analysis have been developed to improve the detection of genes responsible for complex human diseases (as reviewed in Wei et al., 2014). However, models that take multiple SNP markers into account are still not widely adopted and have only recently been applied to plants, including crops, to identify novel candidate genes and gene networks controlling complex agronomic traits. For example, FT is a crucial yet complex trait of interest in barley and other agronomically important crops (Hill and Li, 2016); several studies have reported gene×gene interactions affecting FT in different plant species (Caicedo et al., 2004; Durand et al., 2012; Maurer et al., 2015). Mathew et al. (2018) observed genomic regions with main or higher order epistatic effects overlapping with known candidate genes that were reported previously in barley and closely related species for FT. In sorghum, it is known that Maturity locus 1 (Ma1) represses expression of the floral activator Early heading date 1 (Ehd1), which activates FT to produce florigen for floral induction (Rooney and Aydin, 1999). Li et al. (2018) revealed a significant interaction between the QTL harbouring Ma1 and the QTL harbouring FT through epistasis analysis. The reported gene×gene interactions are consistent with the networking system proposed for the control of the timing of flowering (Blázquez, 2000; Valverde et al., 2004; Imaizumi and Kay, 2006).

To overcome these limitations, gene-set analysis (GSA) has emerged as a more powerful approach than single SNP analysis (Nam et al., 2010). GSA has several advantages. First, GSA can aggregate effects of many SNPs with weak associations. Although individual SNPs may show little or no effect, their interactions may have a non-linear effect if an unbiased analysis for interactions within combinations of SNPs is performed (Wang et al., 2012; Mooney and Wilmot, 2015; Pers, 2016). Secondly, GSA takes allelic heterogeneity into consideration (i.e. different SNPs within a gene linking to a similar phenotype) which is usually not possible in the single SNP GWAS test (Zöllner and Pritchard, 2005; Guan and Stephens, 2011; Jiang et al., 2018). Thirdly, GSA could capture local epistatic interactions between SNPs within a gene and therefore potentially increase prediction accuracies (Zhang et al., 2014; Jiang et al., 2018). As FT is believed to be controlled by a complex interacting gene network probably influenced by the effects of sets of genes (Hill and Li, 2016, and reference therein), testing associations between a phenotype and the cumulative effect of genes may identify more functionally relevant candidate genes with higher accuracy than single SNP GWAS.

Here, we revisited previously published data sets: (i) phenotypic data of 895 barley varieties grown over two years in four different field locations with varying seasonal temperature and rainfall conditions in Western Australia’s South West; and (ii) genotypic data obtained from the targeted resequencing of 174 putative phenology-related genes and gene orthologues (Hill et al., 2019a, b). Building on the previous study, here we aimed to achieve higher statistical power to detect significant genes and gene networks that influence FT in barley by expanding single SNP GWAS analysis to gene-based analysis and epistasis analysis. By taking only SNPs detected within gene-coding regions of putative FT-related genes into account, we first re-calculated the narrow-sense SNP-based heritability of awn emergence as an equivalent to FT (Alqudah and Schnurbusch, 2017). We then re-assessed the association of individual SNPs and FT by standardizing and averaging FT across multiple locations and experimental years. We further grouped SNPs from the same genes into distinct gene sets and tested the association of each gene set with FT. Finally, we identified interacting SNP pairs using a two-way epistasis analysis and determined an expanded and improved gene interaction network which regulates FT in barley.

Materials and methods

Plant material, phenotypic data, and genes enriched in SNPs

Plant material, phenotypic data, and phenology gene-enriched genetic variants were previously reported in detail in Hill et al. (2019a, b). Briefly, 952 barley accessions from 41 countries in Europe, Asia, North and South America, Africa, and Australia were initially selected to represent the global diversity for phenology genes in barley. These accessions represent the entire spectrum of cultivated barley, including two- and six-row genotypes, and winter and spring growth habits. These accessions were grown in 2015 and 2016 at four locations in Western Australia which significantly differ in rainfall and temperature during the growing season. Awn emergence, defined as the number of days from sowing to the first awn emergence above the flag leaf (Z49) (Zadoks et al., 1974), was recorded as an equivalent to FT (Alqudah and Schnurbusch, 2017). A total of 2758 SNPs were enriched from 174 putative genes that are related to phenology and the development of meristem and inflorescences. Full details of field experiments, targeted resequencing of phenology genes, and SNP discovery and filtering were provided in Hill et al. (2019a, b).

Data preparation

The original measurement of days to Z49 for each accession was transformed to standardized FT (FT_D) separately for each growing environment and year using the formula:

$${\displaystyle \begin{array}{l}{\displaystyle \mathrm{F}{\mathrm{T}}_{\mathrm{D}}=\frac{\mathrm{D}\mathrm{a}\mathrm{y}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{Z}{49}_{\mathrm{a}\mathrm{c}\mathrm{c}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}-\mathrm{M}\mathrm{i}\mathrm{n}{(\mathrm{D}\mathrm{a}\mathrm{y}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{Z}49)}_{\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{e}}}{\mathrm{M}\mathrm{a}\mathrm{x}{(\mathrm{D}\mathrm{a}\mathrm{y}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{Z}49)}_{\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{e}}-\mathrm{M}\mathrm{i}\mathrm{n}{(\mathrm{D}\mathrm{a}\mathrm{y}\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{Z}49)}_{\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{e}}}.}\end{array}}$$ (1)

We then averaged FT_D across four locations and two years for each barley variety to minimize the random effect, while not shrinking the genetic effects (Piepho et al., 2008).

Barley accessions or SNP loci with >10% missing data were excluded from analysis. For the remaining missing SNP data in the data set, we inspected each missing datum individually and replaced the missing data manually with the most likely allelic combinations with consideration of linkage equilibrium and allelic state of the individual in other SNP loci. After the filtering, 895 barley accessions and 2758 SNPs remained for heritability estimation and GWAS analysis.

Estimation of narrow-sense SNP-based heritability

A genome-based restricted maximum likelihood method (GREML-LDMS) was used to estimate the heritability of FT using all filtered SNPs. GREML-LDMS corrects linkage disequilibrium biases in the estimated SNP-based heritability (Yang et al., 2015). To calculate narrow-sense heritability from SNP data, h²_SNP, we first computed linkage disequilibrium (LD) scores between SNPs with the block size of 100 kb using the computer software package GCTA (Yang et al., 2011). We used the GREML (a function within GCTA) to estimate the proportion of variance in a phenotype explained by all SNPs (i.e. the SNP-based heritability), following an LD score regression approach as detailed in Yang et al. (2015). h²_SNP was estimated both with and without additional data descriptors (growth habit, row type, and origin of the barley accessions) fitted as fixed effects.

Genome-wide association analysis

We used a linear mixed model (LMM) for GWAS analysis as implemented in the Factored Spectrally Transformed Linear Mixed Models (FaST-LMM) package to perform single SNP, gene-set GWAS, and epistasis analysis (Lippert et al., 2011; Listgarten et al., 2012; Widmer et al., 2014). GWAS are often confounded by population substructure and sample relatedness. LMMs are a powerful and established tool for studying genotype–phenotype relationships. LMMs can capture confounders (e.g. population substructure and family relatedness) of GWAS simultaneously, without requiring prior knowledge of whether the confounders are present or not (Lippert et al., 2011). Its computational efficiency also makes it feasible for an exhaustive search for gene×gene interactions (Lippert et al., 2013; Widmer et al., 2014).

For GWAS analysis, we calculated the first five principal eigenvectors from principal components analysis (PCA) using GCTA (Yang et al., 2011) and subsequently included them as covariates in the model as fixed effects for association analysis. GWAS analysis was conducted using the Python-based program FaST-LMM (Listgarten et al., 2012) following the developers’ instructions (available from http://microsoftgenomics.github.io/FaST-LMM/). Genetic data were formatted into the binary Plink ped input file format (*.bed, *.bim, and *.fam) using Plink 2.0 (Chang et al., 2015). For single SNP association analysis, we used the average FT_D (see Equation 1) of each filtered barley accession as the phenotypic data, all filtered SNPs as genetic data, and the first five principal eigenvectors from the PCA as the covariate. For GSA, we first grouped the SNPs into 174 gene sets with each set of SNPs corresponding to one gene (each gene set had an average of 18 SNPs ranging from 1 to 167). The algorithm as employed in FaST-LMM uses two random effects—one to capture the confounder’s effect and the other to reflect the set association signal—to correct for confounder, and uncovers signal not recoverable by single-SNP GWAS analysis (Listgarten et al., 2013). For epistasis testing, one SNP (the first polymorphic SNP locus) was taken from each gene, as such a filtering approach significantly reduces the required statistical power for multiple testing. The GWAS analysis was then used to test whether pairs of SNPs taken together explain a higher proportion of variance than the sum of the individual effects of each SNP analysed separately (Widmer et al., 2014).

Because the SNPs were enriched from putative genes that were reported to be associated with FT in barley, A. thaliana, and other cereal crops, we adopted a less stringent threshold than Bonferroni correction to define the significance in GWAS. We instead used the Holm’s sequential Bonferroni correction (Holm, 1979) with a significance threshold at P<0.05 to determine significant SNPs, gene sets, and SNP pairs with epistatic interaction. Sequential Bonferroni correction is an adjusted Bonferroni correction depending on rank to maximize the statistical power in GWAS whilst being stringent. ANOVA was implemented using SPSS (Statistical Package for the Social Sciences, SPSS Inc., Chicago, Il, USA) software, and P<0.05 was used as the statistically significant threshold.

Regulatory connections between flowering genes

Interacting network of flowering genes was constructed using STRING, a database of known and predicted gene–gene (protein–protein) interactions (Szklarczyk et al., 2017). In STRING, each protein–protein interaction is assigned a score, as an indicator of confidence of a true interaction. A score of 0.7 was used to assign high confidence when retaining the interaction. Connections between the networks of each key gene were achieved by connecting the overlapping genes and epistatic interactions as revealed in the epistatic analysis.

Results

Flowering time and environmental influence

All 895 barley accessions were grown across multienvironment field trials, conducted over four geographical locations and two years in Western Australia. Significant phenotypic differences of agronomic and phenological traits measured were present for the set of barley genotypes grown in the field at different geographical locations in WA in the 2015 and 2016 growing seasons (Table 1; with more details in Hill et al., 2019a). Average time to flowering for the 895 accessions ranged from 65 d to 85 d, with median time from 72 d to 111 d, after sowing across the trial environments. The range in FTs for all accessions evaluated varied from 42–94 d to 63–136 d across the environments. Geraldton in the North of WA is characterized by a hot and dry environment with a short growing season, with the lowest median number of days to Z49 recorded for any environment (72 d), with a range of 46–89 d recorded in 2015. The trial environments at Esperance (ESP) in Southern WA have a longer, wetter, and cooler growing season, and thus recorded the longest maximum days to Z49 (146 d) in 2016.

Table 1.

Locations and experimental years, major climatic factors, and flowering time mean (days to Z49)

Location (year)	Tmin–Tmax (°C)	Tmean (°C)	Rainfall (mm)	Global solar radiation (MJ m−2)	Growth period (d)	Days to Z49 median (range)
Geraldton (2015)	2–40	17.3	189.8	18.03	182	72 (46–89)
Geraldton (2016)	3–41	15.1	355.4	17.50	210	80 (44–91)
Katanning (2015)	4–36	14.0	550.0	14.63	208	105 (69–132)
Katanning (2016)	-3–38	12.9	256.2	16.87	244	105 (75–131)
Esperance (2015)	1–41	14.7	318.2	12.35	203	104 (60–136)
Esperance (2016)	3–37	13.6	343.6	12.86	201	110 (72–146)
Merredin (2016)	-1–37	14.1	181.4	16.39	191	111 (80–131)

Tmin/Tmax/Tmean: minimum/maximum/mean temperature during the growing season. Environmental data were taken over 200 d since the sowing date during the growth period for comparisons. Modified from Hill et al. (2019a)

The average range to Z49 between the earliest and the latest flowering types was 59 d, showing the considerable genetic difference in controlling the switch from vegetative growth to reproduction among the tested barley accessions. Variation of FT within barley accessions in different environments is strongly influenced by average temperature during the growth period. Growing season average temperature explained 62.9% (P=0.0001) of variance in FT across environments of four locations and two years (Fig. 1), while minimum/maximum temperature, global solar radiation, and rainfall during the growth period had no significant influence on FT (P>0.05). The trial environments received an optimum rainfall throughout the two growing seasons at all four locations.

Fig. 1.

Mean temperature influencing flowering time (days to Z49) in barley in seven experimental environmental sets across four locations in two years. r and p represent correlation coefficient and probability, respectively, assuming a linear relationship between flowering and temperature. Whiskers are standard deviations.

Barley accessions with contrasting growth habit (spring or winter type) had similar standardized days to Z49 (FT_D), as did the barley accessions with different row type (P>0.05). However, average FT_D of barley accessions with different origin was significantly different (ANOVA P<0.0001) (Fig. 2). Barley accessions with different origin also had unequal variances in FT_D (ANOVA, F=9.117, df=44.2, P<0.0001).

Fig. 2.

Phenology of barley accessions with contrasting growth habits, row types, and geographic origin of accessions. Asterisk indicates significant difference in ANOVA. Numbers in parentheses indicate the number of samples, and only the samples positively identified were included.

Gene-set GWAS analysis of flowering time

After filtering (<10% missingness) and pruning to only SNPs located within gene-coding regions of the 174 targeted phenology genes, 895 barley accessions and 2758 SNP markers were retained. Genetic variation of the 2758 SNPs across the barley accessions were not structured by row type, nor by growth habit, nor by geographic origin, confirming previous findings (Hill et al., 2019a, b). Narrow-sense heritability as estimated from all SNP (h²_SNP) was estimated at 0.395±0.048. Specifying the origin of each barley line in the analysis as a fixed effect increased h²_SNP to 0.503±0.056, while including growth habit or row type as fixed effects did not improve the estimation of heritability. Average temperature, as one of the most significant environmental factors, explained 3.8% of the variance in FT_D between barley accessions which was a small yet significant (P<0.001) amount.

Using the sequential Bonferroni correction and significance threshold of P<0.05, GWAS analysis was performed using FaST-LMM (Listgarten et al., 2012), and 170 SNP loci were found to be associated with FT_D across all environments (see the Materials and methods). Systematic biases in GWAS were low, indicated by a λ _GC, the genomic inflation factor, close to 1.1 (Winkler et al., 2014) (Supplementary Fig. S1 at JXB online). These SNPs were located within the gene-coding regions of 32 genes on six chromosomes, with no significant SNPs detected on chromosomes 4H (Fig. 3). The subsequent GSA revealed 113 gene sets, corresponding to 113 putative genes, among the 174 genes that were previously shown as flowering-related genes in cereal crop species or in A. thaliana (Hill et al., 2019a, b), associated with FT_D in the barley accessions. Those significant genes are located on all seven chromosomes (Fig. 4) and from all flowering pathways (Table 2): photoperiod and circadian clock (34 genes), meristem response and development (27 genes), gibberellin signalling and metabolism (19 genes), grain development (15 genes), vernalization regulation (14 genes), and light perception and signalling (10 genes). Among the 170 significant SNPs as detected in single SNP GWAS, 167 SNPs and the 29 corresponding genes they belong to were also detected as part of gene sets to be significantly associated with FT_D (Table 2).

Fig. 3.

Manhattan plot of single SNP GWAS showing significant SNPs that are associated with flowering time in barley accessions. Significant SNPs are shown with larger symbols above the red dashed line as the significance threshold. Significance was determined by sequential Bonferroni correction at P<0.05.

Fig. 4.

Manhattan plot of gene-set GWAS showing significant genes that are associated with phenology in the barley accessions. Significant genes are shown above the red dashed line as the significance threshold as determined by sequential Bonferroni correction at P<0.05.

Table 2.

Genes, their annotation, and associated flowering pathways in barley, as revealed to be significantly associated with flowering time through gene-set analysis

Putative gene name	Annotation (Hv_IBSC_PGSB_r1_HighConf)	Gene ID (Hv_IBSC_PGSB_r1 _HighConf)	Flowering pathway
HvADA2	Transcriptional adapter 2	HORVU5Hr1G095400	Vernalization
HvAGL1	MADS-box transcription factor TaAGL1	HORVU6Hr1G002330	Vernalization and autonomous pathways
HvAGL32	MADS-box transcription factor 31	HORVU2Hr1G098930	Meristem response and development
HvAGLG1	MADS-box transcription factor 34	HORVU5Hr1G095710	Meristem response and development
HvAP2	AP2-like ethylene-responsive transcription factor	HORVU2Hr1G113880	Meristem response and development
HvARF2	auxin response factor 2	HORVU3Hr1G096510	Grain size and reproductive development
HvBB	E3 ubiquitin ligase BIG BROTHER	HORVU4Hr1G055690	Grain development
HvBM1	MADS-box transcription factor 47	HORVU4Hr1G077850	Meristem response and development
HvBM16	MADS-box transcription factor 16	HORVU7Hr1G091210	Meristem response and development
HvBM3	MADS-box transcription factor 18	HORVU0Hr1G003020	Meristem response and development
HvBM5 (HvVRN-H1)	MADS-box transcription factor 14	HORVU5Hr1G095630	Vernalization
HvBM8	MADS-box transcription factor 15	HORVU2Hr1G063800	Meristem response and development
HvBM9	MADS-box transcription factor 7	HORVU7Hr1G054220	Meristem response and development
HvCBF10A	ethylene-responsive element binding factor 13	HORVU5Hr1G080430	Vernalization
HvCBF14	Ethylene-responsive element binding factor 14	HORVU5Hr1G080350	Vernalization
HvCBF2A	Dehydration-responsive element-binding protein 1B	HORVU5Hr1G080310	Vernalization
HvCBF3	C-repeat-binding factor 4	HORVU5Hr1G080420	Vernalization
HvCBF4A	Dehydration-responsive element-binding protein 1B	HORVU5Hr1G080300	Vernalization
HvCBF6	C-repeat-binding factor 4	HORVU5Hr1G080450	Vernalization
HvCBF8A	C-repeat binding factor 3-like protein	HORVU2Hr1G041090	Vernalization
HvCBF9	Dehydration-responsive element-binding protein 1B	HORVU5Hr1G080230	Vernalization
HvCCA1	circadian clock-associated 1	HORVU7Hr1G070870	Photoperiod and circadian clock
HvCDF1	DOF zinc finger protein 1	HORVU2Hr1G017290	Photoperiod and circadian clock
HvCEN	Protein TERMINAL FLOWER 1	HORVU2Hr1G072750	Meristem response and development
HvCIGARP	GRAS family transcription factor	HORVU2Hr1G043780	Gibberellin signalling and metabolism
HvCIGARP-2	SCARECROW-like 1	HORVU3Hr1G091250	Gibberellin signalling and metabolism
HvCK2B	casein kinase II beta subunit 4	HORVU1Hr1G055250	Photoperiod and circadian clock
HvCKX	Cytokinin dehydrogenase 2	HORVU3Hr1G027460	Grain development
HvCMF4	CCT motif family protein	HORVU4Hr1G084020	Photoperiod and circadian clock
HvCMF6b	Zinc finger protein CONSTANS-LIKE 4	HORVU1Hr1G095410	Photoperiod and circadian clock
HvCO11	Zinc finger protein CONSTANS-LIKE 16	HORVU6Hr1G073170	Photoperiod and circadian clock
HvCO2	receptor kinase 3	HORVU6Hr1G072620	Photoperiod and circadian clock
HvCO8	CONSTANS-like 5	HORVU7Hr1G027560	Photoperiod and circadian clock
HvCOP1	Erect panicle 2 protein	HORVU2Hr1G031030	Grain development
HvCry1a	cryptochrome 1	HORVU6Hr1G049950	Light perception and signalling
HvCry2	cryptochrome 2	HORVU6Hr1G058740	Light perception and signalling
HvCYP1	Cytochrome P450 superfamily protein	HORVU2Hr1G081650	Grain development
HvDRF1	Ethylene-responsive transcription factor 4	HORVU1Hr1G060490	Meristem response and development
HvDRF2	Ethylene-responsive transcription factor 4	HORVU6Hr1G050500	Meristem response and development
HvEFS	Histone-lysine N-methyltransferase 2A	HORVU2Hr1G000940	Photoperiod and circadian clock
HvELF3	Early flowering 3	HORVU1Hr1G094980	Photoperiod and circadian clock
HvELF4-like4	ELF4-like 4	HORVU5Hr1G060000	Photoperiod and circadian clock
HvELF7	RNA polymerase II-associated factor 1 homolog	HORVU3Hr1G001430	Photoperiod and circadian clock
HvFCA	FCA-A1	HORVU5Hr1G050820	Photoperiod and circadian clock
HvFD	Lysine-specific histone demethylase 1 homolog 3	HORVU2Hr1G096300	Meristem response and development
HvFT1	FLOWERING LOCUS T 1	HORVU7Hr1G024610	Photoperiod and circadian clock
HvFT2	Protein FLOWERING LOCUS T	HORVU3Hr1G027590	Photoperiod and circadian clock
HvFT3	Protein FLOWERING LOCUS T	HORVU1Hr1G076420	Photoperiod and circadian clock
HvFT5	Protein FLOWERING LOCUS T	HORVU4Hr1G090390	Photoperiod and circadian clock
HvFTL5	Protein FLOWERING LOCUS T	HORVU2Hr1G084540	Photoperiod and circadian clock
HvGA20ox1	gibberellin 20 oxidase 1	HORVU5Hr1G124120	Gibberellin signalling and metabolism
HvGA20ox2	gibberellin 20-oxidase 2	HORVU3Hr1G090980	Gibberellin signalling and metabolism
HvGA20ox2-2	2-oxoglutarate (2OG) and Fe(II)-dependent oxygenase superfamily protein	HORVU1Hr1G070710	Gibberellin signalling and metabolism
HvGA20ox2-2	1-aminocyclopropane-1-carboxylate oxidase 1	HORVU2Hr1G114980	Gibberellin signalling and metabolism
HvGA20ox2-3	2-oxoglutarate (2OG) and Fe(II)-dependent oxygenase superfamily protein	HORVU4Hr1G013840	Gibberellin signalling and metabolism
HvGA20ox3	gibberellin 20 oxidase 2	HORVU3Hr1G089980	Gibberellin signalling and metabolism
HvGA2betadiox7	2-oxoglutarate (2OG) and Fe(II)-dependent oxygenase superfamily protein	HORVU3Hr1G117870	Gibberellin signalling and metabolism
HvGA3ox1	gibberellin 3-oxidase 1	HORVU2Hr1G118350	Gibberellin signalling and metabolism
HvGA3ox2	gibberellin 3-oxidase 2	HORVU3Hr1G022840	Gibberellin signalling and metabolism
HvGARMP	Scarecrow-like transcription factor PAT1	HORVU4Hr1G071670	Gibberellin signalling and metabolism
HvGID1	Gibberellin receptor GID1	HORVU1Hr1G060810	Gibberellin signalling and metabolism
HvGID1L2-3	alpha/beta-Hydrolases superfamily protein	HORVU5Hr1G068140	Gibberellin signalling and metabolism
HvGID1L2-4	alpha/beta-Hydrolases superfamily protein	HORVU5Hr1G069040	Gibberellin signalling and metabolism
HvGID1L2-5	alpha/beta-Hydrolases superfamily protein	HORVU5Hr1G098770	Gibberellin signalling and metabolism
HvGID1L2-8	Acetylesterase	HORVU4Hr1G015550	Gibberellin signalling and metabolism
HvGRP7a	Histone-lysine N-methyltransferase	HORVU4Hr1G003060	Photoperiod and circadian clock
HvGW7	unknown function	HORVU2Hr1G032710	Grain development
HvHYL	alpha/beta-Hydrolases superfamily protein	HORVU0Hr1G004410	Gibberellin signalling and metabolism
HvLFY1	Floricaula/leafy homolog	HORVU2Hr1G102590	Meristem response and development
HvLNG1	unknown function	HORVU2Hr1G063820	Grain development
HvLUX1	Two-component response regulator ARR1	HORVU3Hr1G114970	Circadian clock
HvMADS25-2	MADS-box transcription factor 25	HORVU7Hr1G023940	Meristem response and development
HvMADS25-3	MADS-box transcription factor 25	HORVU7Hr1G024000	Meristem response and development
HvMADS26	MADS-box transcription factor 26	HORVU7Hr1G076310	Meristem response and development
HvMADS68	MADS-box transcription factor family protein	HORVU4Hr1G032440	Meristem response and development
HvMADS75	MADS-box transcription factor family protein	HORVU5Hr1G110470	Meristem response and development
HvNAT	Acyl-CoA N-acyltransferases (NAT) superfamily protein	HORVU7Hr1G113480	Grain development
HvNHL	NHL domain-containing protein	HORVU6Hr1G045970	Grain development
HvPAF	Phytochrome A-associated F-box protein	HORVU1Hr1G058630	Light perception and signalling
HvPFT1	Mediator of RNA polymerase II transcription subunit 25	HORVU5Hr1G054650	Light perception and signalling
HvPhyA	phytochrome A	HORVU4Hr1G008610	Light perception and signalling
HvPhyB	phytochrome B	HORVU4Hr1G053400	Light perception and signalling
HvPhyC	phytochrome C	HORVU5Hr1G095530	Light perception and signalling
HvPI	MADS-box transcription factor 4	HORVU1Hr1G063620	Meristem response and development
HvPI-2	MADS-box transcription factor 2	HORVU3Hr1G091000	Meristem response and development
HvPIF4	Transcription factor EB	HORVU5Hr1G011780	Light perception and signalling
HvPPD-H1	pseudo-response regulator 7	HORVU2Hr1G013400	Photoperiod and circadian clock
HvPRR59	Two-component response regulator-like APRR5	HORVU4Hr1G021010	Photoperiod and circadian clock
HvPRR73	pseudo-response regulator 7	HORVU4Hr1G057550	Photoperiod and circadian clock
HvPRR95	Two-component response regulator-like PRR95	HORVU5Hr1G081620	Photoperiod and circadian clock
HvRLPK	Leucine-rich receptor-like protein kinase family protein	HORVU4Hr1G079040	Grain development
HvRNG	Protein SIP5	HORVU6Hr1G044080	Grain development
HvSCPL33	Carboxypeptidase Y homolog A	HORVU3Hr1G033550	Grain development
HvSHP1	MADS-box transcription factor 13	HORVU1Hr1G023620	Meristem response and development
HvSP1	Protein NRT1/ PTR FAMILY 4.3	HORVU4Hr1G015640	Grain development
HvSPL11	squamosa promoter binding protein-like 2	HORVU6Hr1G031450	Photoperiod and circadian clock
HvSPL12	squamosa promoter-binding-like protein 3	HORVU6Hr1G019700	Photoperiod and circadian clock
HvSPL14	squamosa promoter-binding-like protein 17	HORVU0Hr1G020810	Photoperiod and circadian clock
HvSPL3	squamosa promoter binding protein-like 8	HORVU6Hr1G030490	Meristem response and development
HvSS1	strictosidine synthase-like 3	HORVU5Hr1G091230	Grain development
HvSTK	MADS-box transcription factor 21	HORVU1Hr1G064150	Meristem response and development
HvTEM1	AP2/B3 transcription factor family protein	HORVU3Hr1G010100	Photoperiod and circadian clock
HvTFL1	Protein TERMINAL FLOWER 1	HORVU5Hr1G042230	Meristem response and development
HvTOC1	Two-component response regulator-like PRR1	HORVU6Hr1G057630	Photoperiod and circadian clock
HvTT16	MADS-box transcription factor 29	HORVU6Hr1G032220	Meristem response and development
HvTUBA3	tubulin alpha-4 chain	HORVU4Hr1G009520	Meristem response and development
HvVEL1	Protein VERNALIZATION INSENSITIVE 3	HORVU6Hr1G022770	Vernalization
HvVIN3	Protein VERNALIZATION INSENSITIVE 3	HORVU7Hr1G099250	Vernalization
HvWPSRLK	Mitochondrial transcription termination factor family protein	HORVU2Hr1G061060	Grain development
HvWRKY61	WRKY DNA-binding protein 3	HORVU5Hr1G028340	Grain development
HvZCCTc	Zinc finger protein CONSTANS-LIKE 4	HORVU1Hr1G056120	Vernalization
HvZTLa	Kelch repeat-containing F-box family protein	HORVU7Hr1G099010	Photoperiod and circadian clock
HvZTLb	Adagio-like protein 1	HORVU6Hr1G022330	Photoperiod and circadian clock
a HvAG1	MADS-box transcription factor 3	HORVU3Hr1G026650	Meristem response and development
a HvCK2A	Protein kinase superfamily protein	HORVU0Hr1G030500	Photoperiod and circadian clock
a HvPAP2	Auxin-responsive protein IAA17	HORVU3Hr1G031460	Light perception and signalling
b HvBM7	MADS-box transcription factor 1	HORVU4Hr1G067680	Meristem response and development
b HvCO1	B-Box-type zinc finger transcription factor	HORVU7Hr1G043030	Photoperiod and circadian clock
b HvCry1b	cryptochrome 1	HORVU2Hr1G079220	Light perception and signalling
b HvEDL2	EID1-like F-box protein 2	HORVU2Hr1G034270	Photoperiod and circadian clock
b HvGA2ox3	gibberellin 2-oxidase	HORVU3Hr1G072810	Gibberellin signalling and metabolism

Significance was determined by sequential Bonferroni correction (P<0.05). The detailed list with chromosome position is in table S1 in Hill et al. (2019a). Annotation and Gene ID follows Hv_IBSC_PGSB_r1_HighConf.

^a Significant only in single SNP GWAS analysis.

^b Significant only in epistasis analysis.

Epistatic effects of genes associated with flowering time

Two-way (interaction of two SNPs) epistasis analysis revealed 19 pairs of SNPs (sequential Bonferroni corrected P<0.05), among the overall 30 276 pairs between each of the phenology genes studied here, interacting to influence FT. Depending on the combination of SNPs in their allelic state, 12 pairs significantly promoted earlier flowering (–8 d), and seven pairs were linked with later flowering (+10 d) when compared with average FT (Table 3). A homozygote at an alternative state (‘GG’ versus ‘TT’ in the reference genome) in HvELF7 (an RNA polymerase II-associated factor 1 homologue gene) interacted with six other SNPs promoting earlier flowering, while the HvGA2ox3 (a gibberellin 2-oxidase gene) homozygote at an alternative state (‘AA’ versus ‘GG’ in the reference genome) interacts with other genes to delay flowering in barley (Fig. 5). For example, cultivar ‘UWA2Rsel9506’ which has genotype ‘GG’ in HvELF7 tends to flower earlier when HvCO1 (a zinc finger protein CONSTANS-LIKE gene) has genotype ‘GG’ across all experimental locations. Seven accessions (‘07T741’, ‘B559’, ‘B751’, ‘Han 85-222’, ‘I92-562’, ‘ICB104039’, and ‘Lao Wu Hu Xu Mai’) with HvGA2ox having genotype ‘AA’ and HvCKX (a cytokinin dehydrogenase gene) having genotype ‘GG’ usually flower later across our trials. When homozygous in an alternative state (‘GG’ versus ‘CC’ in the reference genome), HvPhyB (a phytochrome B gene) interacts with two other genes (HvNHL, an NHL domain-containing protein gene, and HvTOC, a two-component response regulator-like PRR1 gene) to promote early flowering, while when in the heterozygous state, this gene interacts with other genes (HvSPL3, a squamosa promoter-binding protein-like gene) to promote late flowering. Eight out of the 13 genes revealed to have epistatic interactions were also significant in the SNP-set GWAS analysis, while the remaining five were defined as insignificant both in the single SNP and gene-set GWAS analyses (Table 2).

Table 3.

SNP–SNP interaction in determining flowering time in barley as revealed by epistasis analysis

Gene_1	FTD	Gene_2	FTD	Gene interaction	FTD
Gene interactions to promote early flowering
HvCBF8A (CC)	0.64±0.12	HvELF7 (GG)	0.52±0.09	CC–GG	0.48±0.09
HvCO1 (GG)	0.63±0.12	HvELF7 (GG)	0.52±0.09	GG–GG	0.48±0.10
HvCry1b (TT)	0.64±0.12	HvELF7 (GG)	0.52±0.09	TT–GG	0.48±0.09
HvBM7 (CC)	0.64±0.12	HvELF7 (GG)	0.52±0.09	CC–GG	0.48±0.09
HvPhyB (CC)	0.65±0.12	HvELF7 (GG)	0.52±0.09	CC–GG	0.48±0.09
HvFT1 (CC)	0.64±0.13	HvELF7 (GG)	0.52±0.09	CC–GG	0.48±0.09
HvCK2B (GG)	0.64±0.12	HvCO1 (CC)	0.64±0.12	GG–CC	0.55±0.13
HvCO1 (GG)	0.63±0.12	HvZCCTc (CC)	0.60±0.13	GG–CC	0.58±0.11
HvPhyB (GG)	0.60±0.13	HvNHL (GG)	0.61±0.12	GG–GG	0.56±0.08
HvPhyB (GG)	0.60±0.13	HvTOC1 (TT)	0.62±0.13	GG–TT	0.56±0.08
HvPhyA (AA)	0.61±0.13	HvZTLa (GG)	0.64±0.12	AA–GG	0.57±0.13
HvCBF8A (TT)	0.59±0.12	HvEDL2 (TT)	0.64±0.13	TT–TT	0.58±0.10
Gene interactions to delay flowering
HvPhyA (GG)	0.64±0.12	HvZTLb (GG)	0.65±0.13	GG–GG	0.66±0.11
HvPhyB (CC)	0.64±0.12	HvSPL3 (CC)	0.61±0.12	CC–CC	0.69±0.11
HvSLN1 (CC)	0.67±0.12	HvCO8 (TT)	0.70±0.16	CC–TT	0.74±0.14
HvCKX (CC)	0.70±0.15	HvGA2ox3 (AA)	0.69±0.15	CC–AA	0.77±0.15
HvFT2 (GG)	0.70±0.16	HvGA2ox3 (AA)	0.69±0.15	GG–AA	0.77±0.15
HvFT2 (GG)	0.70±0.16	HvCBF6 (TT)	0.69±0.15	GG–TT	0.77±0.15
HvCBF6 (TT)	0.69±0.15	HvCKX (CC)	0.70±0.15	TT–CC	0.77±0.15

Flowering time (days to Z49) was standardized to 0–1 as FT_D (see the Materials and methods). Letters in parentheses indicate the genotype of the first SNP of the gene. FT_D is presented as mean ±SD. Note that the average FT_D across all samples was 0.64±0.12

Fig. 5.

Significant flowering genes and their regulatory connections in barley (Hordeum vulgare L.). Putative gene name and gene IDs were from Ensembl Plants Hordeum vulgare Genome assembly 082214v1 that was archived in STRING (Szklarczyk et al., 2017). The interactions, including type and effects, were based direct (physical) and indirect (functional) associations from computational prediction and knowledge transfer between organisms, as implemented in STRING (Szklarczyk et al., 2017).

Gene interaction network in regulation of flowering time

Using key genes involved in flowering regulation in barley as recorded in the comprehensive protein–protein interaction database ‘STRING’ and also including additional candidate genes as revealed in our gene–gene interactions (epistatic interaction) analysis, we constructed a complex gene regulatory network (Fig. 5). The network involved 18 genes that were identified as significant in the above gene-based associated analysis. These genes are known to have roles in light signalling (e.g. HvPhyC), photoperiod response (HvPpD-H1), circadian clock (HvELF3), and development of the inflorescence meristem (HvCEN). Twenty-one genes were uncharacterized in the Hordeum vulgare genome assembly 082214v1.

Discussion

We have previously identified 429 functional alleles within the coding regions of 95 genes associated with FT in barley using single-marker GWAS (Hill et al., 2019a). In this study, by expanding to GSA and epistasis analysis, we achieved higher statistical power, and with potentially high accuracy, to detect significant genes and gene networks that influence FT in barley. We have identified 121 genes that have been associated with FT in barley, including 26 that have not been described in barley in previous research. All 121 genes have been previously described in dicot A. thaliana, and monocot cereal crops (e.g. rice, maize, and sorghum), indicating that many of the flowering genes are conserved across angiosperms including dicots and monocots (Blümel et al., 2015). FT genes involved in the photoperiod, vernalization, circadian clock, and gibberellin biosynthesis pathways were previously studied in barley (Turner et al., 2005; Wang et al., 2010; Maurer et al., 2015; Mathew et al., 2018). Our GSA detected essential genes involved in the key flowering pathways and confirmed that these genes were indeed controlling the FT in the barley accessions with a broad geographic origin. We note that our SNPs have been enriched from putative flowering genes; it is highly likely that there are additional genes, and gene interactions between flowering genes and other genes that may not directly be involved in flowering, influencing flowering in barley. Further research into the genetic mechanism of flowering in barley should expand to include genome-wide genetic variants.

Our gene-set association analysis detected key photoperiod response genes controlling FT. The photoperiod response gene Photoperiod 1 (Ppd-H1), located at chromosome 2H, is a pseudoresponse regulator gene. This gene has previously been identified as one essential gene for providing adaptation to photoperiod in barley by flowering induction under long days (Turner et al., 2005). It is known that the Ppd-H1 dominant allele induces early flowering in wild and winter barley varieties, while recessive ppd-H1 delays flowering in spring barleys (Turner et al., 2005; Jones et al., 2008). The second photoperiod gene Ppd-H2, also known as HvFT3 in barley, located on chromosome 1H, was shown to regulate FT under short days (Börner et al., 2002; Wang et al., 2010). GSA identified HvCEN as a significant flowering gene, corroborating the report from Comadran et al. (2012). TFL1, the homologue of HvCEN, is a key regulator of FT by controlling the development of the inflorescence meristem in A. thaliana (Hanano and Goto, 2011). HvCEN and associated QTLs were also reported to be associated with components of grain yield traits in barley (Comadran et al., 2012; Pasam and Sharma, 2014; Sharma et al., 2018). Saade et al. (2016) reported that the HvCEN locus promoted early FT, and resulted in higher grain yield, under salt stress conditions.

Among the three light receptor phytochrome genes—HvPhyA, HvPhyB, and HvPhyC—identified as associated with FT in our GSA, HvPhyC has previously been reported as an essential component in photoperiodic flowering in barley (Faure et al., 2012; Nishida et al., 2013; Pankin et al., 2014; Hill et al., 2019a). As phytochromes are involved in plants’ ability to intercept and translate light signals, they play a crucial role in modulating and regulating growth and development (Mathews, 2010). The HvPHYC gene was reported to interact with several other photoperiod response genes under different photoperiods (Pankin et al., 2014). Meanwhile, existing evidence suggests that variation at the HvPHYC locus has no pleiotropic effects on important agronomic traits and starch pasting properties (Nishida et al., 2013; Pankin et al., 2014). As such, Ibrahim et al. (2018) suggested that HvPHYC can be used effectively in barley breeding programmes to manipulate FT for yield improvement for varieties in stressful growing conditions.

Circadian clock-controlled mechanisms enable plants to measure changes of photoperiod as a cue for seasonal changes in their environment and therefore control developmental transitions, such as from vegetative growth to initiating flowering (Shim et al., 2017). Previous reports identified HvELF3 as one of the key genes affecting the circadian clock (Faure et al., 2012; Zakhrabekova et al., 2012), which was also confirmed in this study. The HvELF3 locus regulates flowering under the influence of photoperiod (Boden et al., 2014). In A. thaliana, it is known that ELF3, LUX, and ELF4 form a protein complex, termed the evening complex (EC). This complex represses the expression of PRR9 and LUX (two core circadian components in A. thaliana) through binding to LUX-binding sites (reviewed in Shim et al., 2017). Huang et al. (2016) recently reported that the PhyB–ELF3 complex forms one of the signalling hubs that connects red light signalling with the circadian clock. It is not clear whether the circadian clock-controlled mechanisms involving ELF3, ELF4, LUX, PRR9, and PhyB operate in the same way in barley as in A. thaliana. However, HvELF3, HvELF4, HvLUX, HvPRR9, and HvPhyB were all identified as significant in controlling FT in our gene-set test.

It is known that the early flowering of some barley genotypes is closely linked to gibberellin biosynthesis (Boden et al., 2014). We identified 19 genes related to gibberellin biosynthesis [e.g. HvGA20ox1 (GA20 oxidase 1)] as significant flowering genes in the barley accessions we investigated. Our findings corroborate with the notion that gibberellin is an important signal in flower development in barley. In A. thaliana, paclobutrazol—a gibberellin biosynthesis inhibitor—significantly reduces the long hypocotyl and petiole phenotypes of Arabidopsis elf3 mutants (Filo et al., 2015). As discussed above, ELF3 is a key gene in a tripartite transcriptional complex, the EC. Filo et al. (2015) further suggested that the role of the EC in the regulation of gibberellin biosynthesis and flowering in dicots is shared with monocots and is a highly conserved mechanism for growth control. As such, mechanisms of the circadian clock-controlled pathway linking regulation of gibberellin biosynthesis and flowering induction, as reported in A. thaliana, may provide a useful template for exploring clock-controlled mechanisms in barley.

Fourteen genes that were reportedly involved in the vernalization pathway have been identified in the GSA. The interaction of Vrn-H1, Vrn-H2, and Vrn-H3 has been reported as an important mechanism controlling flowering in response to vernalization in barley (von Zitzewitz et al., 2005). HvBM5 (equivalent to HvVrn-H1), a MADS-box transcription factor gene, was identified as a significant flowering gene, and was also previously reported to promote the transition from the vegetative to the reproductive phase (Hemming et al., 2008). In the interaction, HvVrn-H1 represses the expression of Vrn-H2 (a zinc-finger CONSTANS); in turn, that represses Vrn-H3 in regulating flowering as the response to vernalization (Yan et al., 2003, 2004). Vrn-H3 (equivalent to HvFT1) in barley was thought to be a central integrator of different FT pathways (Yan et al., 2006). Yan et al. (2006) also reported that the Vrn-H3 gene in both barley and wheat is responsible for natural allelic variation in vernalization requirement. Five FT genes (HvFT1, HvFT2, HvFT4, HvFT5, and HvFTL5) were identified as significantly influencing FT in this study. These genes were observed to play different roles in their response to photoperiod, while HvFT1 has an essential role in the transition from the vegetative growth to reproductive stage (Alqudah et al., 2014).

We identified 22 genes involving 19 two-way epistatic interactions in either promoting early flowering or delaying flowering. Epistatic interactions have previously been reported in barley. Yan et al. (2004, 2006) have previously reported significant two-way epistasis between vernalization genes VRN-H1 (syn. HvBM5) and VRN-H3 (syn. HvFT1), and between Vrn-H1 and Vrn-H2, to play an essential role in FT regulation in barley. Griffiths et al. (2003) postulated that FT genes HvGI, Vrn-H2, Vrn-H1, and HvCO1 could be involved in two-way epistatic interactions. Cuesta-Marcos et al. (2010) proposed that Vrn-H1 (HvBM5), Vrn-H2, Vrn-H3 (HvFT1), and Vrn4 could interact to determine vernalization sensitivity in barley. A few of the epistatic interactions revealed in this study could be linked to previously reported interaction in barley or A. thaliana. For example, the interaction of homozygous HvFT2 and HvGA2ox3 delayed flowering in our study, which is consistent with the previous report by Filo et al. (2015). Our results also demonstrate the extensive epistatic interactions controlling the FT between genes involved in response to photoperiod, circadian clock pathway genes, response to vernalization, and gibberellin biosynthesis (Fig. 5). HvELF7, a homologue of the RNA polymerase II-associated factor 1 gene, is notable. This gene interacted with six other genes involved with photoperiod and vernalization to induce flowering up to 10 d earlier. Its effects were consistent across our experimental locations and years, implying that its role is probably independent of environmental impacts.

HvCO1 is another key gene identified in this study. HvCO1 and HvCO8 were involved in four epistatic interactions in influencing flowering. It is known that CONSTANS (CO) plays a crucial role in the photoperiodic regulation of flowering in A. thaliana (Kim et al., 2008). At least eight homologues of CO-like genes (HvCO1–HvCO8) were identified in barley, but their roles in controlling the FT pathway are not clear (Griffiths et al., 2003; Cockram et al., 2012). Our findings for HvCO1 and other genes involved in photoperiodic regulation and vernalization could provide some testing hypothesis of the role of CO in the regulation of flowering in barley.

Interestingly, HvCO1 was involved in epistatic interactions promoting early flowering, while HvCO8 interacting with HvSLN1 delayed flowering in the studied barley accessions, implying the possible different roles that different homologues of CO-like genes may play in the regulation of FT in barley. The broad epistatic interactions in the regulation of FT in barley as revealed in our study suggest the presence of other functional networks of genes involved in controlling FT. Based on the fact that more genes and their interactions were identified as important in regulating FT in barley, this study added more details to the gene regulatory network that Hill and Li (2016) proposed. Our results on epistatic interaction and proposed gene regulatory networks could provide further insight to refine the current model of the regulatory network controlling flowering in barley and other cereal crops (e.g. Woods et al., 2017), while further studies, such as with knock-out accessions, may validate the observed interaction effects and regulatory network.

The main environmental factors that influence FT include the ambient temperature and day length. In sorghum, temperature explained 69.4% of the variation in average FT in different environments (Li et al., 2018). Similarly, our study found that 62.9% of the variation in FT of a barley line was due to variation in average temperature in the growing locations. Gene and environment interactions explained 3.85% (P<0.001) of variance for FT. This figure, although much less than that by the average temperature, was found to be highly significant. FT in barley is highly heritable. Broad-sense heritability of FT was estimated at 88% in wild barley (Herzig et al., 2018). In maize, the genetic architecture of FT is predominantly determined by small additive loci with few environmental interactions, and FT is also highly heritable (h₂ >0.85) (Buckler et al., 2009). Our estimate of heritability from 2758 SNPs was 0.503 if the origin of the experimental accessions was included as a fixed effect, while it was only 0.395 if the origin was not specified. Previously, we reported that peak SNPs at the identified loci explained 31–78% of the phenotypic variance for phenology in different environments (Hill et al., 2019a).

Both our current and previous estimates of heritability seem to be low, which could be explained by four aspects. (i) There may be more genes that are essential parts of the network regulating FT in barley yet to be captured in our study. For example, Bouché et al. (2016) curated a database containing 306 genes that were reported to have functions and interactions within the flowering pathways in A. thaliana, while we analysed 174 putative genes. (ii) Causal SNPs related to FT could be located far from the known gene in its regulatory regions; therefore, SNP enrichment based on genes could fail to capture the effect. (iii) The epistatic effect could be more extensive because of the existence of a complex regulation network in controlling flowering. (iv) Broader sampling to include samples from broader genetic background and origin could be required. Future research that builds on the insights generated from this study, and with the aim of finding the missing heritability and the genes that are important in regulating FT in barley, will help to decipher the genetic mechanism of flowering regulation, and therefore facilitate barley breeding programmes to increase performance and grain yield under optimal cultivation conditions as well as under stress.

GWAS has been a powerful tool to connect genomic variation (SNPs) to complex phenotype, while pinpointing the actual genes underlying biology is still not straightforward. Our previous research (Hill et al., 2019a) demonstrated that targeted enrichment of SNPs from function-related genes combined with GWAS could provide great opportunities to associate DNA variations with complex phenotypes in plants. In this study, we further demonstrated that GSA could provide higher power to detect genetic association than the analysis of SNPs individually. We suggest that GSA is particularly useful for dissecting the genetic determinants of complex traits such as FT, as it is likely that many SNPs with small effects contribute to these complex traits, while their effects are difficult to detect when testing SNPs individually (Holmans, 2010). Our research also shows that the incorporation of analysis of gene interaction and gene-set GWAS offers great promise in the characterization of the biological pathway of genetic determination of complex traits. It should be noted that, despite the power to connect sequence diversity to complex traits, GSA has its limits. First, GWAS analysis so far revealed that most of the significant SNPs fall within the category of non-protein coding, and many are a distance away from the known gene (Maurano et al., 2012); it is not clear how far the flanking sequencing of each gene should be included in the mapping of SNPs to a gene set (Fridley and Biernacka, 2011). Further, as with the single SNP GWAS analysis, GSA reveals the genetic changes to be correlated with a particular phenotype; this does not mean that genes identified by the studies control the phenotype, which needs to be tested in controlled experiments.

Supplementary data

Supplementary data are available at JXB online.

Fig. S1. Q–Q plots showing the low level of systematic biases in genome-wide association study (GWAS) results.

Author contributions

CL, DM, and PT conceived the project; CH and ZX generated the genotypic data; TA, CL, and SW generated the phenotypic data; TH performed the data analysis; KC performed spatial analysis, and TH and CH completed the manuscript with inputs from CL, PT, TA, and the other co-authors.