Genetic control and phenotypic characterization of panicle architecture and grain yield-related traits in foxtail millet (Setaria italica)

Multi-environment QTL mapping identified 23 stable loci and 34 co-located QTL clusters for panicle architecture and grain yield-related traits, which provide a genetic basis for foxtail millet yield improvement. Panicle architecture and grain weight, both of which are influenced by genetic and environmental factors, have significant effects on grain yield potential. Here, we used a recombinant inbred line (RIL) population of 333 lines of foxtail millet, which were grown in 13 trials with varying environmental conditions, to identify quantitative trait loci (QTL) controlling nine agronomic traits related to panicle architecture and grain yield. We found that panicle weight, grain weight per panicle, panicle length, panicle diameter, and panicle exsertion length varied across different geographical locations. QTL mapping revealed 159 QTL for nine traits. Of the 159 QTL, 34 were identified in 2 to 12 environments, suggesting that the genetic control of panicle architecture in foxtail millet is sensitive to photoperiod and/or other environmental factors. Eighty-eight QTL controlling different traits formed 34 co-located QTL clusters, including the triple QTL cluster qPD9.2/qPL9.5/qPEL9.3, which was detected 23 times in 13 environments. Several candidate genes, including Seita.2G388700, Seita.3G136000, Seita.4G185300, Seita.5G241500, Seita.5G243100, Seita.9G281300, and Seita.9G342700, were identified in the genomic intervals of multi-environmental QTL or co-located QTL clusters. Using available phenotypic and genotype data, we conducted haplotype analysis for Seita.2G002300 and Seita.9G064000,which showed high correlations with panicle weight and panicle exsertion length, respectively. These results not only provided a basis for further fine mapping, functional studies and marker-assisted selection of traits related to panicle architecture in foxtail millet, but also provide information for comparative genomics analyses of cereal crops.


Introduction
Foxtail millet (Setaria italica) is one of the most important cereal crops that domesticated in China about 10,000 years ago (Hu et al. 2018;Lu et al. 2009). The mature panicle of foxtail millet contains many primary branches that are attached to the main axis (often referred to as the rachis), several secondary branches on the primary branches, tertiary branches on the secondary branches, and each of the tertiary branches bears numbers of spikelets (grains). A major goal of foxtail millet breeding is to improve the grain yield per unit growing area by cultivating varieties with large panicles, long branches, a high grain number, and enlarged grains. Despite a significant progress has been made in foxtail millet grain production during 40 years of scientific breeding in China (Diao et al. 2014), the molecular and genetic mechanisms underlying foxtail millet grain yield, especially panicle architecture, remain unclear.
Most of foxtail millet cultivars have only one tiller, which bears one panicle (Doust 2007). Traits related to panicle architecture, such as panicle length (PL), panicle diameter (PD), primary branch number (PBN), primary branch length (PBL), and grain number per panicle (GNP), mainly determine the grain yield per plant. Panicle architecture is mainly determined by the fate of the meristem, and by the timing of the meristem phase shift from the branch meristem to the spikelet meristem (Kyozuka et al. 2014). Studies on rice and maize have shown that many genes involved in CLAVATA -WUS signaling pathway, such as FLORAL ORGAN NUMBER 1 (FON1) (Suzaki et al. 2004), FON4 (Chu et al. 2006), OsWUS, ZmWUS, ZmWUS2 (Nardmann and Werr 2007), and thick tassel dwarf 1 (TD1) (Bommert et al. 2005), and many MADSbox transcription factors, including OsMADS34 (Gao et al. 2010), OsMADS14, OsMADS15, andOsMADS18 (Kobayashi et al. 2012), affect panicle architecture by regulating meristem size and specification of meristem identity. Additionally, the MADS-box transcription factors can also regulate inflorescence branching by repressing the expression of REDUCED CULM NUMBER 4 (RCN4), a homolog of TERMINAL FLOWER 1 (TFL1)/CENTRORADIALIS in rice . Overexpression of TFL1 homologs (RCN1 and RCN2) in rice (Nakagawa et al. 2002) and CENTRORADIALIS 1 (ZCN1) to ZCN6 in maize (Danilevskaya et al. 2010) delay the changes from branch shoot to floral meristem and lead to a highly branched inflorescence. Gain of function mutations in DENSE AND ERECT PANICLE1 (DEP1) enhances meristematic activity, resulting in shortening of the inflorescence internode, and increased number of grains per panicle (Huang et al. 2009a). In addition, mutants of DEP2 and DEP3 also exhibit a characteristic erect panicle phenotype and increased panicle length and grain size (Qiao et al. 2011;Zhu et al. 2010). Phytohormones, including auxin, cytokinin, and gibberellin, play essential roles in regulating inflorescence meristem identity, initiation, and enlargement. Loss-of-function mutations in genes that participate in local auxin biosynthesis, signaling and transport significantly affect panicle architecture development (Komatsu et al. 2001(Komatsu et al. , 2003Phillips et al. 2011). Grain number 1a (Gn1a) encodes a cytokinin oxidase/dehydrogenase (OsCKX2) that degrades cytokinin, loss-of-function mutations in Gn1a accumulate higher levels of cytokinin in inflorescence meristems, resulting in a larger number of branches and spikelets (Ashikari et al. 2005). By contrast, the dysfunction mutant of LONELY GUY (LOG), which encodes an enzyme catalyzing the conversion of inactive cytokinin nucleotides to the active free-base forms, displays a small inflorescence with a decreased number of branches and spikelets (Kurakawa et al. 2007). Class 1 KNOTTED 1-like homeobox (KNOX) genes, such as rice homeobox 1 (OSH1) (Tsuda et al. 2011), maize knotted 1 (KN1) (Vollbrecht et al. 2000), and Arabidopsis SHOOT MERISTEMLESS (STM) (Long et al. 1996), play a central role in promoting shoot apical meristem identity by decreasing the levels of GA and increasing the amount of cytokinin. Due to a KNOX-mediated transcriptional feedback loop, overexpression of Grain Number per Panicle1 (GNP1), which encodes rice GA20ox1 that degrades active GAs, increases grain number and yield by increasing cytokinin activity in rice panicle meristems (Wu et al. 2016).
Most research on genes related to panicle architecture has been conducted using maize and rice. In contrast, only a few genes related to panicle architecture have been cloned from Seteria mutants and molecularly characterized in detail. By screening for visible inflorescence mutant phenotypes via an N-nitroso-N-methylurea (NMU) mutagenesis of Setaria viridis, Huang et al. (2017) identified two sparse panicle mutants, spp1 and spp3. Both were found to have disruptive mutations in the SvAUX1 (AUXIN1) gene. Further study revealed that loss-of-function mutations in SvAUX1 and ZmAUX1 disrupt the inflorescence branch development in S. viridis and maize, leading to sparse panicle phenotypes. In previous studies, we isoloated two mutants with abnormal panicle architecture from the EMS mutant library constructed in our laboratory. One is the loose-panicle mutant, in which a single G-to-A transition in the fifth intron of the WRKY transcription factor gene results in three disorganized splicing events in mutant plants, leading to a lax primary branching pattern and aberrant branch morphology (Xiang et al. 2017). The other is simads34, in which an alternative splicing event introduces an early termination codon in SiMADS34. This results in increased panicle width, primary branch length, and number of primary branches, but 1 3 decreased panicle length and grain weight in simads34 compared with wild-type plants (Hussin et al. 2021).
Forward genetic studies have identified many quantitative trait loci (QTL) associated with panicle architecture or grain yield from various bi-parental populations. Doust et al. (2005) conducted QTL analyses on the basis of differences in the inflorescence between foxtail millet and green foxtail. They detected 14 replicated QTL for PBN and primary branch density, spikelet number, and bristle number. Using an F 2 population derived from different foxtail millet cultivars, Fang et al. (2016) identified 12 QTL for PL, PD, panicle weight per panicle (PW), grain weight per panicle (GWP), and 1000-grain weight (TGW). Wang et al. (2017) Jaiswal et al. (2019) performed a genome wide association study (GWAS) on 10 agronomic traits using 142 foxtail millet accessions, and identified 17 and 10 loci for grain yield and TGW, respectively. However, neither of those QTL have been cloned, nor have any of the candidate genes located in the QTL intervals been isolated. This has hindered the understanding of the mechanisms underlying foxtail millet panicle architecture and grain size development.
In the present study, an RIL population was developed from derivatives of the cross between the foxtail millet cultivars Ai 88 and Liaogu 1. Panicle architecture varied greatly between the two parents and within the RIL population. A large-scale and multi-environment analysis using the RIL population was carried out. An ultra-high density genetic map was constructed to explore the genetic control of panicle architecture and yield-related agronomic traits in 13 environments. The QTL mapping identified 159 QTL, whose genetic intervals contain many candidate genes involved in panicle development. The favorable QTL alleles from either parent will be of great value to optimize panicle architecture and increase the grain yield of foxtail millet.

Plant materials
A foxtail millet RIL population comprising of 333 lines was used in this study. This RIL population was generated from a cross between a backbone line Ai88 and an elite variety Liaogu1 as described by He et al. (2021). From 2015 to 2018, the RIL population was grown at seven geographical locations during the growing season ( Fig. 1 and Table S1). Nanbin Farm (NB,109.19°E/18 (Fig. 1a). Plants were grown in 3 m × 0.9 m plots in three lines using standard agronomic practices (e.g., irrigation, weeding, and pest control). Five individuals in the middle of each row were harvested individually to score the traits.

Phenotype evaluation
When panicles and grains were fully mature, PL was measured for the main panicle, and PD was measured at the thickest location on the main panicle. The PEL was measured from the uppermost node to the panicle base. Harvested panicles were air-dried and stored at room temperature for 1 month, then primary branches were removed from the panicle and grains were removed from the branches for measurements. The PBL was measured using a ruler. The PW, GWP, and TGW were evaluated using a Mettler-Toledo analytical balance (model XP204S/M, Mettler-Toledo, Greisensee, Switzerland). The PBN per panicle and grain number per primary branch (GNB) were counted manually. Both PL and PEL were measured in all 13 environments; PD, PW, and GWP were evaluated in 12, 11, and 10 environments, respectively; PBN and TGW were measured in two environments; and PBL and GNB were measured in one environment (Fig. 1). All traits were measured with three to five replicates.

Statistical analysis of panicle phenotypic variations
Analyses of all the phenotypic variations in the RIL population, including calculations of the mean value, standard deviation, skewness, kurtosis, as well as the broad-sense heritability (h 2 ) and correlation analysis, were performed using R packages. Analysis of variance (ANOVA) was carried out to test the statistical significance of differences in traits of RILs among various environments. Variance components of genotype (V G ), genotype and location interaction (V G×L ), genotype and year interaction variance (V G×Y ), and residual variance (V r ) were estimated using a mixed linear model using the R package lme4. Broad-sense heritability was calculated using the following formula: where n L and n Y are the number of location and year, respectively.

QTL mapping, QTL comparison and candidate gene identification
The high-density genetic map constructed by He et al. (2021) was used in this study. The R/qtl package was used to perform QTL mapping using the CIM model with a scanning window size of 5 cM. The loci with LOD (logarithm of odds ratio) over 2.5 were considered as QTL and the confidence intervals were estimated using the 1.5 LOD-drop method (He et al. 2021). By running 1,000 permutation tests for each trait, high-confidence QTL (those with LOD over the significance thresholds, p < 0.05) were identified. Meanwhile, QTL across different environments for the same trait Primary branch length, h Primary branch number, i Grain number per branch, and j 1000-grain weight variations of RILs in different locations. Analysis of variance (ANOVA) was used to determine the significance of differences in traits among RILs grown in diverse environments. Different letters indicate significant difference among different environments. Statistical analysis was performed using ANOVA with Tukey's test were considered to be the same when the supporting intervals overlapped and the additive effects originated from the same parental line. Overlapping of genomic regions for QTL controlling different traits was indicative of co-located QTL clusters. QTL and environment interaction study followed the methods described by Li et al. (2015). First, the loci with LOD over 4.7 (the average threshold (LOD ≥ 4.7) of all traits by 1,000 permutation test without environment interaction) was considered as a QTL for Q × E interaction. Many QTL exceeded the threshold (LOD ≥ 4.7) for some traits such as PL, PEL, and PD. Therefore, to reduce the false positive rate, we considered the top 20 QTL for traits with more than 20 QTL with LOD ≥ 4.7.
The QTL nomenclature followed the rules described by Mccouch et al. (1997), and QTL on the same chromosome were listed in alphabetical order. The QTL with a positive or negative additive effect for a specific trait indicate that the increase in the phenotypic value of the trait is contributed by the alleles from Liaogu1 or Ai88, respectively. For QTL comparison, if the physical coordinates of a reported QTL were within a 0.5-Mb interval of a QTL detected in this study, they were considered to be the same QTL.
Based on amino acid similarity, candidate genes located in QTL intervals showing homology to genes related to panicle architecture in rice were identified. Amino acid similarity was further confirmed using tools at Phytozome (https:// phyto zome. jgi. doe. gov/ pz/ portal. html). Candidate genes with nucleotide variations in the untranslated region (UTR) or coding sequence between the two parental lines were subjected to haplotype analysis. Protein domain enrichment analysis was performed using STRING software (https:// string-db. org/ cgi/ organ isms).

Haplotype analysis of candidate genes
We collected the available data for phenotypes related to panicle architecture for 900 accessions in our previous study (Jia et al. 2013). The genotype data for those accessions were obtained by high-depth resequencing (unpublished). The haplotype analysis was performed using in-house python and R scripts.

Phenotypic variation and broad-sense heritability
In this study, all nine traits exhibited diverse phenotypic variations and obvious transgressive segregations in the RIL population ( Figure S1 and Table S1). All phenotypes in the RIL population showed normal distributions, suggesting that these traits related to panicle architecture were controlled by QTL. The performances of PW, GWP, PL, and PD were influenced by the geographical location (Fig. 1b-e and Table S1). The phenotypic variations in the RIL population were quite stable across years at a given location, except for PD and PEL in 2017_CY and 2018_CY. The average phenotypic values for PW, GWP, PL, and PD in the RIL population increased dramatically from the lower latitude locations to the higher latitude locations. For example, the RIL popula-  Table S1). Similar to PL, the traits PW, GWP, and PD showed phenotypic variations that were highly correlated with latitude (or photoperiod) (Fig. 1b, c, e), while the variation in PEL was not correlated with latitude (Fig. 1f).
In this study, most of the traits related to panicle architecture displayed significant variations in different environments. Therefore, we evaluated the broad-sense heritability (h 2 ) of PW, GWP, PL, PD, and PEL on the basis of RIL phenotypic data in six to seven locations across 3 to 4 years ( Table 1). The PL trait exhibited the highest heritability (h 2 = 0.93), while GWP showed the lowest heritability

Correlation analysis
We investigated the relationships between pairs of traits related to panicle architecture in each of the 13 environments ( Figure S1). The traits PL, PD, and PEL were evaluated in all 13 environments, except for PD in 2016_CZ. The trait PL was positively correlated with PD and PEL in seven and nine environments (p < 0.05), respectively. We detected positive correlations between PD and PW and between PL and GWP in all environments tested (p < 0.01). The traits PW and GWP were evaluated in 11 and 10 environments, respectively, and PW was highly positively correlated with GWP in all environments tested (p < 0.001). We analyzed the correlations of PEL and PW, PEL, and GWP in 11 and 10 environments, respectively, and found that PEL was positively correlated with PW and GWP in 2016_NB, 2017_ZZ, 2018_ ZZ, and 2016_GZL, and PEL was negatively correlated with PW in 2017_CY (p < 0.05). We measured the TGW and GNB in 2017_TY, and found that TGW was positively correlated with PL, PD, PW, and GWP (p < 0.05); and GNB was positively correlated with PD, PW, and GWP (p < 0.05).
The PBN was positively correlated with PL in 2017_TY and 2018_CY (p < 0.01), and positively correlated with PW and GWP in 2017_TY (p < 0.001). The PBL was positively correlated with PL, PD, PW, and GWP in 2018_CY (p < 0.05), but negatively correlated with PEL in 2018_CY (p < 0.05). No significant correlations were detected for other traits related to panicle architecture in this study. Overall, PL, PD, TGW, GNB, PBN, and PBL were positively correlated with PW and GWP; and PEL was negatively correlated with PD in most environments investigated.

QTL mapping
In total, 239 loci, including 73 high-confidence loci, formed 159 QTL for nine traits that were detected under 13 environments across 4 years ( Table S2). The LOD values of these QTL ranged from 2.51 to 22.36, and explained 0.29% to 25.55% of the phenotypic variations. Out of the 159 QTL, 34 were identified in two to 12 environments (Table S2).
We also compared the genomic intervals of QTL controlling different traits, and found 17, 14, one, and two genomic intervals shared by two, three, four, and five QTL for different traits, respectively (Table S3 and Figure S2). The Q × E analysis showed that more than 67% (89) of the top 20 QTL detected in interaction studies overlapped with non-Q × E interaction QTL (Table S4). Forty-three major effect Q × E QTL were not detected in the non-Q × E interaction study for PW, GWP, PD, GNB (Table S4).

Panicle weight
Twenty-four QTL related to PW were detected in 11 environments across 3 years, explaining 0.58% to 8.69% of the phenotypic variation (Table S2)

Grain weight per panicle
Sixteen QTL for GWP were detected across nine environments, and explained 2.70%-7.58% of the phenotypic variation. One of them, qGWP4.1 was identified in two environments. The favorable allele was from Ai88. The remaining 15 QTL were only identified in a single environment. The additive effects of seven QTL were derived from Liaogu1.

Panicle length
Thirty-five QTL mapped on all chromosomes were found to be related to PL in 13 environments, accounting for 0.29%-25.55% of the phenotypic variation. Of them, qPL7.2 and qPL9.5 were identified across 12 and seven environments, respectively, and the additive effect of these two QTL was contributed by Liaogu1. qPL9.5 accounted for 14.95%-25.55% of the phenotypic variation. qPL5.2 was identified in three environments, and the favorable allele originated from Liaogu1. qPL2. 6,qPL3.1,qPL3.8,qPL4.3,qPL5.1,qPL7.3,qPL9.2,and qPL9.3 were detected in two environments, all of the additive effects for PL were derived from Liaogu1, except for qPL3.8 and qPL5.1, which were derived from Ai88. The other 24 QTL for PL were identified in a single environment. The additive effects of 12 of them were from Liaogu1 and the others were from Ai88.

Panicle diameter
Twenty-seven QTL for PD located on all chromosomes except chromosomes 4 and 7 were detected in 12 environments across 4 years, and accounted for 0.
Only one QTL for GNB was detected in 2017_TY, accounting for 3.47% of the phenotypic variation. The additive effect of qGNB9 was from Liaogu1.

Haplotype analysis of candidate genes
Analyses of sequencing coverage and previously collected phenotypic data identified two candidate genes (Seita.2G002300 and Seita.9G064000) with sufficient phenotypic data for haplotype analysis (Fig. 3). For Seita.2G0023000, there were 11 haplotypes in 603 accessions (Fig. 3a). H5 had the lowest panicle weight, and the PW of H5 was significantly different from those of H1, H2, H3, H6, and H7 (Fig. 3b). Most of the variations in H5 were located in the 5' untranslated region (UTR), suggesting that 5'UTR of Seita.2G002300 might play an important role in panicle weight regulation. There were 12 haplotypes of the candidate gene of qPEL9.1 (Seita.9G064000) in 318 accessions. The PEL of H1 was significantly lower than those of H2 and H3, indicating that Seita.9G064000 was a potential candidate gene controlling the PEL of foxtail millet (Fig. 3c, d).

Discussion
Foxtail millet has many excellent characteristics as a model system for C 4 plants, because of its small diploid genome, short growth duration, self-fertility, fertile seed setting, small morphological stature, and easy management in the laboratory (Doust et al. 2009). In the present study, we investigated the phenotypic variations in a RIL population of

G T A C G C G C C C A A A C G T A T A T T T C T
H10 H11

A T T G G T A A C C T T G C G A T C C G T T C C
C H7 N N N N N N N   27.0 ± 6.9

G T T G G T A A C C C T C G A T A T T T C T A G C G C T T T C T A G C G C T T T C T A G T G T T T T C T A G C G C T T T C T
H12 27.4 ± 1.3

Comparison the QTL identified in this study with previous studies
Based on the physical coordinates of the QTL confidence intervals, we compared the genomic regions of the QTL identified in this study with those detected in other bi-parental and natural populations. Eleven QTL overlapped with QTL previously detected in other studies. One QTL was 0.5 Mb away from a previously identified QTL (Table S2). The genomic regions of qPW2.6, qPD8.2, and qPEL5.6 overlapped with those of qPW2, qPD8-2, and qNL5 identified in a F 2 population of foxtail millet comprising 543 lines derived from a cross between Aininghuang and Jingu 21 (Wang et al. 2019 (Fang et al. 2016). The genomic regions of six QTL were the same as those of six QTL identified in a previous study from a natural population of 916 accessions (Jia et al. 2013 (Jia et al. 2013).

Candidate genes located at QTL related to panicle architecture or grain yield
The inflorescence of foxtail millet, like all panicoid grasses, is a compound raceme called a panicle. The inflorescence meristem is composed of a rachis, primary branches, secondary branch meristems, and tertiary branches with a number of spikelet meristems, which develop into two-flowered spikelets. Many genes involved in specifying inflorescence meristem identity have been cloned. In rice, ASP1 encodes a TOPLESS-related transcriptional co-repressor that is involved in the regulation of meristem fate. A recessive aps1 mutant displays a disorganized branching pattern and aberrant spikelet morphology . APO1 temporally regulates meristem identity in rice. The inflorescence meristem of apo1 is converted into a spikelet meristem, and produces a small number of primary branch meristems, resulting in small panicles (Ikeda et al. 2005). OsMADS14, OsMADS15, and OsMADS18 are APETALA1 (AP1)/FRUITFULL (FUL)-like genes involved in inducing the transition from the shoot apical meristem to the inflorescence meristem in rice. Triple knock-down plants in the pap2 mutant show inhibited transition of the meristem to the inflorescence meristem (Kobayashi et al. 2012). In this study, we identified four candidate genes homologous to ASP1, APO1, OsMADS15, and OsMADS18 of rice, in the genomic regions of qPD6.2, qGWP4.1/qPL4.3/qPW4.2, qPW2.1, and qPW2.6/qTGW2.3, respectively. These genes may play essential roles in regulating PW, GWP, TGW, and PL in foxtail millet. Moreover, delays in spikelet meristem specification lead to iterations of branching, resulting in larger panicles that could potentially produce more grain.
In the dominant gain-of-function mutant tawawa1-D, the activity of the inflorescence meristem is extended and spikelet specification is delayed, resulting in prolonged branch formation and increased numbers of spikelets. In contrast, a reduction in TAW AWA 1 expression by RNAi results in a similar but stronger small inflorescence phenotype in which both primary and secondary branches are reduced (Yoshida et al. 2013). We identified a gene (Seita.9G222400) homologous to rice TAW AWA 1 located at qPBN9.2, a QTL related to PBN, suggesting that this gene may play a role in the formation of primary branch number in foxtail millet.
OsMADS22 and OsMADS56 are homologs of Arabidopsis SVP and SOC1, respectively, and regulate inflorescence branching by repressing the expression of RCN genes in rice . The gene SiMADS56 (Seita.9G342700) was identified at the genomic region of qPL9.6/qPD9.2/ qPEL9.3, which was detected 23 times in 13 environments for three traits. Thus, SiMADS56 may play an essential role in regulating branching in foxtail millet, independently of the environment.

Phytohormones might play important roles in foxtail millet panicle architecture and grain yield
Auxin plays a key role in determining axillary meristem initiation and outgrowth. Two of the candidate genes detected in this study, Seita.5G243100 and Seita.7G126900, showed homology to the auxin biosynthesis genes YUC1 and TDD1 in rice. Seita.1G317400, Seita.3G136000, Seita.4G101300, and Seita.5G241500 were predicted to encode auxin efflux carrier proteins involved in auxin transport. Cytokinin and GAs play antagonistic roles in regulating reproductive meristem activity. Increased cytokinin activity leads to higher grain number, whereas GAs negatively affect meristem activity. In the genomic regions of qPBN5/qPEL5.2, we identified a gene (Seita.5G140300) showing homology to the rice gene Gn1a, which encodes a cytokinin oxidase/dehydrogenase (OsCKX2) that degrades cytokinin (Ashikari et al. 2005). Intriguingly, we also detected a rice DST homolog (Seita.9G064000) located at qPEL9.1. In rice, DST enhances grain production through controlling Gn1a/OsCKX2 expression. Seita.9G004400, located at the genomic region of qPBN9.1, showed homology to rice GNP1, which encodes a GA20ox1 protein. KNOX proteins function as modulators, and balance cytokinin and GA activity in the meristem. Increased expression of the GA catabolism genes GA2oxs in NIL-GNP1 TQ decreases GA accumulation, resulting in increased cytokinin activity, which consequently improves grain number and yield (Wu et al. 2016).

Conclusion
In summary, we analyzed the phenotypic variations in nine traits related to panicle architecture in one to 13 environments, and found that phenotypic variations in the RIL population varied across different geographical locations. The QTL mapping revealed 239 loci, including 73 high-confidence loci, forming 159 QTL for nine traits. Of these, 34 QTL were identified in two to 12 environments, and 34 were pleiotropic QTL related to two to five traits. We anticipate that further analyses of these QTL will provide a foundation for further genetic improvement of the yield of foxtail millet.