Introduction

Cotton (Gossypium spp.) is an important crop that provides most natural fiber for industrial textiles. Gossypium hirsutum L., also called upland cotton, is the most widely cultivated cotton species and accounts for more than 90% of worldwide production because of its high yield, wide adaptability, and acceptable fiber quality (Zhang et al. 2008; Page et al. 2013). Cotton fibers are single-celled trichomes that provide a unique experimental system to study cell development processes, such as cell growth, biosynthesis, and response mechanisms for outside signals (Qin and Zhu 2011; Haigler et al. 2012; Walford et al. 2012; Guo et al. 2016). Cotton fiber has been involved in pioneering research concerning the genetic basis of its yield and quality through forward or reverse genetics (Haigler et al. 2012; Said et al. 2015).

Recently, molecular selection strategies based on quantitative trait loci (QTL) mapping are efficient tools for breeders to combine economically important traits to create a superior cultivar (Xu et al. 2012; Smykal et al. 2016). Identifying the QTL and dissecting the genetic basis of agronomy traits are critical steps and prerequisites for marker-assisted breeding in crops. Linkage mapping and linkage disequilibrium (LD) mapping are the mainstream methods for dissecting the genetic architecture of complex traits in crops (Mackay and Powell 2007). Traditionally, bi-parental populations, such as F2, recombinant inbred lines (RILs) and chromosome segment substitution lines (CSSLs), are utilized to excavate the QTLs or genes and to evaluate the biological effects for plant phenotypes (Xu et al. 2012; Said et al. 2015). For qualitative traits or main effect loci, linkage mapping is an efficient way to determine the target genes (Han et al. 2015; Ma et al. 2016; Liu et al. 2016a). However, the resolution of quantitative trait mapping in bi-parental populations is usually low (Mackay and Powell 2007). Association mapping offers a high resolution with either prior information on candidate genes or a genome scan with very high marker coverage (Zhao et al. 2011; Li et al. 2013b). Compared to linkage mapping, association analysis is more efficient and powerful for the dissection of complex traits (Cavanagh et al. 2008; Atwell et al. 2010). In addition, association analyses have been successfully employed in many crops (Zhao et al. 2011; Li et al. 2013b; Xu et al. 2016; Sun et al. 2017).

In most cases, researchers conduct an association analysis based on a natural population from a diverse collection of inbred lines or germplasms that may constitute a group with population structure related to geographical origin or reproductive isolation. The population structure will disrupt the association mapping and give false-positive results. The multi-parent advanced generation inter-cross (MAGIC) population has richer genetic diversity and recombination rate without population structure, which provides power and resolution for gene targeting (Cavanagh et al. 2008; Huang et al. 2015; Pascual et al. 2016). The MAGIC population is suitable for genetic map developing and QTL mapping (Huang et al. 2011, 2015; Bandillo et al. 2013). Recently, MAGIC populations have been created for genetic analysis in crops, such as rice (Bandillo et al. 2013), wheat (Huang et al. 2012; Delhaize et al. 2015), maize (Dell’Acqua et al. 2015), chickpea (Agarwal et al. 2015), and cotton (Islam et al. 2016).

Fiber quality traits are quantitative traits with a complex genetic basis (Fang et al. 2017). Several fiber characteristics, such as fiber length (FL), fiber strength (FS), micronaire value (MV), fiber elongation (FE), fiber uniformity (FU), and short fiber content (SFC) are the main determining factors for fiber quality. Dissection of QTLs for these fiber traits will help to enhance the genetic architecture of fiber quality. Recently, several studies were focused on the genetic bases of fiber quality traits using simple sequence repeat (SSR) or single-nucleotide polymorphism (SNP) markers (Said et al. 2015; Nie et al. 2016; Sun et al. 2017; Huang et al. 2017). In this study, an upland cotton MAGIC population containing 960 lines was developed by a convergent cross based on 8 parents. An association analysis was performed based on phenotypic data from five environments, and genotypic data were generated from 284 polymorphic SSRs. This study aimed to take advantage of the MAGIC population to explore the genetic architecture of 6 fiber traits and to provide molecular genetic basis for fiber improvement in breeding.

Materials and methods

Development of the upland cotton MAGIC population

The upland cotton MAGIC population was developed with eight parents. The parents of MAGIC (PMs) were diverse accessions with abundant phenotypic characteristics for high yield and fiber quality (GY2, GY4, GY5, and GY6), pest resistance (KC9, CQ2, and CQ13) and Fusarium wilt and Verticillium wilt resistance (KB10) (Table S1). They were widely distributed in the Yangtze River valley and Yellow River valley and were provided by CCRI (Cotton Research Institute of Chinese Academy of Agricultural Sciences), HAAFS (Hebei Academy of Agriculture and Forestry Sciences), ACYU (Agricultural College of Yangtze University), HSMS (Hubei Province Seed Management Station), and TCCSF (Taicang City Cotton Seed Farm in Jiangsu province) (Table S1).

The convergent crossing followed the development of an eight-way cross or funnel crossing (Cavanagh et al. 2008; Huang et al. 2015), and the MAGIC population was obtained by inter-crossing eight parents with three generations and successive selfing (Fig. 1). In the first stage, multiple parents (G0) were inter-mated to generate four bi-parental crosses (two-way): GY5 × KB10 (AB), GY2 × KC9 (CD), CQ13 × GY6 (EF), and CQ2 × GY4 (GH) at the Agricultural College of Yangtze University, Jinzhou, Hubei province in the summer of 2004. Then, four F1 plants (G1) were inter-crossed to derive four-way crosses (G2), including AB × CD (ABCD), and EF × GH (EFGH) at Sanya in Hainan province in the winter of 2004. The four-way crosses formed a female parent group (ABCD) containing 200 individuals and a male parent group (EFGH) containing 200 individuals for subsequent crossing. The last stage involved individual inter-crossing between female and male parent groups that formed 200 inter-mated hybridizations to derive the eight-way crosses (G3) EFGH × ABCD (ABCDEFGH) in 2005. Only 168 of 200 mated generations were successfully multiplied.

Fig. 1
figure 1

Schema for MAGIC population development based on eight founders. G0–Gn were generations of crossing and selfing. The parents, A–G, were GY5, KB10, GY2, KC9, CQ13, GY6, CQ2, and GY4, respectively

Full size image

To create homozygous individuals, the eight-way cross generations (G3) were self-pollinated with multiple generations. In the first self-crossing, the 168 G3 heterozygotes were self-crossed and generated ten plants from each line in 2006. Approximately 1680 filial lines were obtained after expanding propagation. However, some of the lines were lost in the process of selfing through single seed descent. Finally, a total of 960 lines were preserved after selfing for more than 5 generations from 2007 to 2011. The 8 parents (PMs) and 960 MAGIC lines (MLs) formed the population panel for this study.

Fiber quality phenotyping

The fiber quality traits were measured under multiple environments. The phenotypic data of eight PMs and 960 MLs were collected from five natural environments at Huanggang, Hubei province (HG), Zhoukou, Henan province (ZK), and Jingzhou, Hubei province (JZ). The five natural environments were defined as E1 (HG, 2013), E2 (HG, 2014), E3 (ZK, 2014), E4 (ZK, 2015), and E5 (JZ, 2015). The 968 lines were grown during the planting season at each location. The field plots followed a completely randomized block design. Each plot contained one row of MLs and multi-rows of PMs. The row length and spacing were set as 5 and 1 m, respectively. Each row included 10 plants with 40 cm between individuals. Fiber samples were collected from 20 normally opened bolls from the middle fruit branches of plants in each row. The fiber quality traits, including the fiber upper half mean length (FUHML, mm), fiber strength (FS, cN/tex), micronaire value (MV), fiber elongation (FE, %), fiber uniformity (FU, %), and short fiber content (SFC, %), were measured by HVI1000 Automatic Fiber Determination System in Institute of Cotton Research, Shihezi Academy of Agricultural Sciences. FE data in 2014 Huanggang (E2) and Zhoukou (E3) were missing and only obtained from three environments. The remaining five traits were obtained from all five environments.

Phenotypic data analysis

The analysis of variance (ANOVA), correlation, and repeatability analysis for fiber quality traits were conducted using IBM SPSS Statistics 21 (SPSS, Chicago, IL, USA). Best linear unbiased predictions (BLUP) were used to estimate phenotypic traits across multiple environments based on a linear model. The phenotypic data from BLUP and the broad-sense heritability (h2) were calculated with the R software (http://www.r-project.org). The statistical analysis of the mean, standard error (SE), and Kendall’s tau-b correlation coefficient was calculated using IBM SPSS Statistics 21 (SPSS, Chicago, IL, USA). The ANOVA was carried out to evaluate the effects of genotype (G), environment (E), and the interactions between genotype and environment (G × E).

SSR genotyping

Five young fresh leaves were picked from five plants of each line. The genomic DNA of 8 PMs and 960 MLs was extracted using the modified CTAB method (Paterson et al. 1993). A 10-µL reaction volume for the SSR reaction system was performed for PCR amplification (Lin et al. 2005). The Fragment Analyzer™ 96-Capillary Automated CE System and gel electrophoresis (6% denaturing polyacrylamide gel for SSR and 8% non-denatured polyacrylamide gel electrophoresis for SSCP-SSR, silver staining developing after electrophoresis) were used to visualize the PCR products. A total of 3871 SSR primers were selected for detecting polymorphisms from the interspecific (Gossypium hirsutum × Gossypium barbadense) genetic map, which contained 5152 loci and spanned 4696.03 cM (Li et al. 2016b). After screening the SSR primers by eight PMs, the polymorphic markers were used to genotype the 960 lines (Table 2). The 5 cM range nearly covered by SSR markers were set as effectively covered fragments. The adjacent markers with gaps ≤ 10 cM were treated as the same covered fragments.

Population structure, kinship and LD decay analysis

The genotypic data of polymorphic markers were used to calculate the genetic eigen value of the MAGIC population. Population structure and kinship could cause false positives or deviations in association mapping. Principal component analysis (PCA) and cluster analysis were used to confirm a reasonable population structure. PCA was conducted by the Powermarker software 3.25 (Liu and Muse 2005) and Ntsys 2.1 (Adams and Rohlf 2000). The relative kinship coefficients (K) were estimated by SPAGeDi version 1.4b (Hardy and Vekemans 2002). The PCA matrix for phenotypic variations was calculated by IBM SPSS Statistics 21, whereas the contribution rate of the K matrix was calculated with TASSEL 3.0 (Bradbury et al. 2007). The linkage disequilibrium (D′, r2 and p value) was calculated with TASSEL 3.0. The parameter r2 was used to graphically represent the LD curves with the R software.

Marker–trait association

Marker–trait association analyses for the six fiber quality traits based on the interspecific genetic map (Li et al. 2016b) were conducted using a general linear model (GLM) and mixed linear model (MLM) with the TASSEL 3.0 software package. The naive GLM model ignores the characteristic vector from the population structure. The MLM-incorporated kinship (K matrix) was used as the concomitant variant to eliminate random effects. The threshold for the significance of associations between SSR markers and traits was set as p < 0.01 (− log10(p) > 2), and a Bonferroni threshold was set, so that p < 0.00352 (p = 1/n, − log10(1/284) = 2.45), which has been widely adopted in the previous studies (Wang et al. 2012; Liu et al. 2015b, 2016b). The sequences of significant associated markers were searched from CottonGen Database (http://www.cottongen.org) and assigned a genome location (NAU-genome database of TM-1, Zhang et al. 2015) using a BLAST (Altschul et al. 1994) search with E ≤ 1e−10.

Results

Upland cotton MAGIC population development and phenotypic characteristics of fiber quality

Eight upland cotton accessions with good characteristics in different traits were selected as the founders (PMs) to develop the MAGIC population (Table S1). An eight-way inter-cross was designed to aggregate the alleles from eight PMs. Overall, 960 stable inbred MAGIC lines (MLs) were obtained after selfing for more than six generations from eight-way lines (Fig. 1). Eight PMs and 960 MLs were used for phenotyping, genotyping, and the association analysis.

The data of five fiber quality traits, including FUHML, FS, MV, FU, and SFC, were collected from E1 (HG, 2013), E2 (HG, 2014), E3 (ZK, 2014), E4 (ZK, 2015), and E5 (JZ, 2015), while the FE data were only collected from E1, E4, and E5. These traits showed rich variations in different environments (Table S2). The coefficients of variation (CV) for six traits ranged from 1.17% (FU-E2) to 20.13% (SFC-E3) in MLs and ranged from 0.72% (FU-E3) to 16.58% (FE-E4) in PMs (Table S2). The six traits, including FUHML, FS, FE, FU, MV, and SFC, varied from 25.11 to 31.13 mm, 25.85 to 39.18 cN/tex, 2.86 to 9.66%, 82.46 to 87.27%, 4.36 to 5.67, and 7.01 to 13.16% in PMs, respectively. The variations of the six traits were wider in MLs than in PMs and varied from 22.34 to 33.36 mm, 12.04 to 43.72 cN/tex, 1.80 to 10.62%, 78.44 to 88.29%, 2.77 to 6.74, and 4.20 to 23.80%, respectively (Table S2).

The six traits were also merged with BLUP to eliminate the environmental impact. The results also showed that phenotypic variations in MLs were more abundant than variations in PMs. The CVs of six traits in BLUP were 2.12, 3.87, 0.65, 0.10, 3.48, and 1.93% in PMs, whereas they were 2.56, 4.03, 10.60, 0.26, 3.87, and 3.26 in MLs. The broad-sense heritability (h2) of six traits ranged from 40 to 81% in MLs. Only the FU and SFC had relatively low h2, which were 42 and 40%, respectively (Table 1). The ANOVA showed that the genotype (G) and environment (E) had significant effects (p < 0.001) on fiber quality traits (Table S3). The results suggested that both genotype and environment affected phenotypic variation, and genotype played a stable role in most traits. Correlation analysis showed that FUHML and FS were significantly correlated with other traits (p < 0.01) (Table S4). Significant positive correlations were found among six paired traits, FUHML-FS, FUHML-FU, FS-FU, FS-MV, FE-MV, and FU-MV, and there were significant negative correlations in the remaining pairs, except for FE-FU and FE-SFC.

Table 1 Statistics description and comparison of phenotypic variations between PMs and MLs
Full size table

Genotyping and genetic diversity

A total of 644 (17.15%) polymorphic markers screened from 3871 SSR markers by 8 PMs were used to genotype the 960 MLs. Among them, only 284 markers showed clear electrophoresis bands and considerable polymorphism, and the minor allele frequency (MAF) was > 0.05. The 284 polymorphic SSR markers were distributed across 26 chromosomes from 3 (Chr04) to 16 (Chr05 and Chr13) (Table 2). The effective fragment covered by SSR markers accounted for approximately 42.98% of the whole genome and ranged from 19.67% (Chr04) to 65.03% (Chr16). The mean values of genetic diversity for the PMs and MLs were 0.415 and 0.465, respectively (Table 2). In addition, the average polymorphism information content (PIC) values of PMs and MLs were 0.346 and 0.390, respectively. These results suggested that the genetic diversity of MLs based on SSR markers was higher than the diversity of PMs.

Table 2 Summary of polymorphic SSR markers and comparison of the gene diversity and PIC between PMs and MLs
Full size table

Population kinship, genetic structure and linkage disequilibrium

The relative kinship coefficients were calculated as the pairwise relatedness between MLs. The results showed that the kinship values after standardized treatment were low with an average value of 0.057 in MLs, but the values were higher with an average value of 0.490 in PMs. In addition, 52.3% of pairwise kinship coefficients among MLs were 0, and approximately 97.0% of pairwise coefficients were less than 0.3 (Fig. S1). The kinship matrix (K matrix) was used to correct random error in the marker–trait association. The K-matrix could explain 24.81, 20.60, 9.21, 11.41 12.73, and 15.28% of the phenotypic variation for FUHML, FS, FE, FU, MV, and SFC, respectively (Table S5).

PCA was used to reflect the genetic divergence for MLs and PMs. The top three principal components (PCs), PC1, PC2, and PC3, accounted for 28.76, 13.06, and 7.99% of the genetic variations, respectively. The PC1–PC2 plots showed a dispersive distribution without obvious clusters (Fig. S2). In addition, the PC1–PC2 plots of MLs were scattered more widely than the PMs. It was suggested that the MLs obtained more diversity without population structure. In addition, for the PCA matrix, the top three PCs explained 14.2, 16.8, 0.2, 9.9, 7.7, and 9.8% of the phenotypic variations for FUHML, FS, FE, FU, MV, and SFC, respectively. It was obviously smaller than the K matrix.

The squared Pearson correlation coefficients (r2) were measured by 284 SSR markers to describe the LD relationship between markers. The LD showed a very low level in MLs (Fig. 2). A total of 40,186 possible pair combinations were obtained with an average of r2 = 0.029. The distribution of r2 values was concentrated in the < 0.1 interval that accounted for 93.6% of all pairwise coefficients and 49.1% of the coefficients were < 0.01. In addition, 27.02, 33.70, and 43.49% of the coefficients (r2) were revealed to be significant at p < 0.0001, p < 0.001 and p < 0.01, respectively. The LD decay rate was measured by an r2 plot based on the corresponding genetic distance across the entire genome. The average distance of pair combinations was ~ 0.60 cM with the r2 plot dropping to 0.2, and it was ~ 0.76 cM with the r2 plot dropping to 0.1 (Fig. 2c). The 0.76 cM approximately equaled 400 kb in physical genome map.

Fig. 2
figure 2

Summary of LD in 960 MLs. a Distribution of r2 and p values for 284 SSR markers on 26 chromosomes. b Histograph of r2. The column diagram indicates the percentage of corresponding intervals for r2. The curve shows cumulative distribution for r2. c The LD decay curve was determined according to r2

Full size image

Association mapping QTLs for fiber quality traits

The MAGIC population was calculated with a low population structure. The naive general linear model (GLM) and mixed linear model with K matrix (MLM (K)) were used to execute the marker–trait associations based on the phenotypic data of six fiber traits. The Q–Q plot of naive GLM deviated materially from the predicted line, which suggested that there was a higher rate of false positives in GLM. Compared to the GLM, the MLM model could correct the error from the random effects (Fig. S3). For this study, the MLM (K) was chosen as the optimal model.

Two thresholds, p < 0.01 (− log10(p) = 2) and a Bonferroni threshold of p < 0.00352 (1/284, − log10(p) = 2.45) were set as the significance levels. When using the traits data from the BLUP results and MLM (K) model, a total of 52 marker–trait associated loci were significantly associated with 6 traits at p < 0.01 (Table 3), while 31 of the 52 loci were identified with p < 0.00352 (Table S6). The FUHML was associated with the maximum loci of 18, while the other 5 traits, including FS, FE, FU, MV, and SFC, identified 12, 2, 4, 7, and 9 loci, respectively (p < 0.01, Table S6). The allele frequency of the significantly associated markers was dependable as the MAF > 7.15% (Table 3). The percentage of phenotypic variation explained by SSR markers (R2) ranged from 0.71 to 5.47% (Table S6). Furthermore, among the 52 loci, 47 loci shared a significant association with the results using a naive GLM model (Table S7). In addition, 134 significant (p < 0.01) marker–trait association loci were found using trait data from individual environments (E1–E5) (Table S8). A total of 35 significant loci associated with the BLUP data were identified in individual environments. For example, MON_SHIN0376b, which was associated with MV based on BLUP data, were detected in E3, E4, and E5.

Table 3 Significant SSR markers associated with six fiber traits after BLUP
Full size table

The 52 marker–trait associated loci were from 40 SSR markers and distributed on 22 chromosomes, except for Chr03, Chr10, Chr17, and Chr22 (Fig. 3). Based on the marker sequences and the NAU-genome database for TM-1, 37 of the 40 identified SSR markers were located on a physical map using BLAST methods (Table S9). The other three markers were below the threshold or located in uncertain positions. There were nine markers associated with multiple traits (Table S10). For example, MON_CGR5525 was associated with three traits, including FUHML, FU, and SFC. Based on studies in recent years, 14 markers were found to be mapped with the same traits or related traits (Fig. 3, Table S11). For example, BNL1231, which was associated with FS, FE, and SFC, was also located in the loci controlling FUHML, FS, MV, and some yield or plant type traits in other studies.

Fig. 3
figure 3

Distribution of the marker–trait loci identified by association analysis on the published genetic map. Genetic distances in centimorgans (cM) are indicated on the left side of the linkage map. The marker name and associated fiber quality traits are indicated on the right side. Asterisk indicates that the markers were the overlapped marker in this study with other studies, as listed in Table S11

Full size image

The candidate regions for marker–trait loci were set around the LD decay distance as 400 kb (r2 < 0.1). The numbers of genes ranged from 4 to 107 with an average of 53 in the candidate regions (Table S9). There were three regions covering only few genes that were less than ten. They were TMB0029b, MON_DPL0906, and MON_DPL0375c, which were significantly associated with the same trait FUHML, linked 2, 3, and 4 genes in the confidence genome intervals (Tables S9 and S12). In addition, there were 20 genes with the functions mainly affected fiber development via resource references, located in neighboring site of the significant SSR loci by BLAST (Table S13). Six genes, Gh_A01G1843 (GhPIP1-2), Gh_A01G1915 (GhACT1), Gh_A09G1461 (GhACT1), Gh_A13G0162 (GhLIM1), Gh_D01G0018 (GhPIP2-4), and Gh_D01G1810 (GhCER6) were located in the candidate regions with a short distance < 400 kb. In addition, the other 14 genes were near the marker–trait loci within < 3 Mb. Several reported genes were found multi-copy loci to be related to different loci; for example, GhACT1 was linked with four loci, HAU4552b, MON_CGR5707, MON_DPL0042, and NAU7153.

Discussion

In this study, an association analysis was performed to explore the genetic basis of fiber quality traits in an upland cotton MAGIC population. Association mapping is a powerful and popular tool for studying the genetic basis of complex traits in plants (Yu et al. 2008; Ingvarsson and Street 2011; Yang et al. 2012). The efficiency of association analysis is determined by the diversity of a population, the experimental population size, and reasonable statistical methods (Yan et al. 2009; Bush and Moore 2012). We developed a MAGIC population with sufficient diversity and individuals. The MAGIC population takes advantage of populations that are known to have high efficiency, high precision, and low false-positive rate (Mackay and Powell 2007; Cavanagh et al. 2008; Huang et al. 2015).

The wider range of phenotypic and genotypic variations generated in MAGIC recombination lines is beneficial for carrying out genetic research of complex traits (Cavanagh et al. 2008; Huang et al. 2015). To maintain relatively high levels of diversity in the MAGIC population, eight parents were selected from different areas with various characteristics and convergent crosses were performed to obtain sufficient recombination (Fig. 1, Table S1). Multiple parents could carry more alleles than a bi-parental program. Although the eight parents had relatively low polymorphism at the genetic level, they created more abundant phenotypic and genotypic variations in the MAGIC lines. The screened SSR markers used for genotyping demonstrated that the MAGIC population had higher genetic diversity compared to the PMs (Table 2) as well as higher diversity than the population used in the association analysis for cotton (Zhao et al. 2014; Liu et al. 2015a). In addition, the MAGIC population had relatively abundant phenotypic variations compared to the PMs. The multiple parents inter-cross is a breeding method that could gather favorable alleles and create excellent new varieties through recombination (Cavanagh et al. 2008; Huang et al. 2015). In this study, more MLs had better properties in terms of fiber quality with fiber lengths > 30 mm and fiber strengths > 30 cN/tex. The selection of germplasm from the cotton MAGIC population could potentially improve fiber traits. There are 960 MLs in the MAGIC population, which is larger than the natural population used in association analysis of upland cotton conducted in the previous studies (Islam et al. 2016; Nie et al. 2016; Sun et al. 2017), which ensures a high efficiency in QTL mapping.

The LD decay is an important factor that determines the mapping resolution of an association analysis. The LD level is affected by recombination and some factors that generate change of allele frequency distribution, such as mutations, gene drift, selection, and migration. Recombination is the major factor that decides the speed of LD decay. On the other hand, the LD can reflect the historical condition of recombination in most inbred plant populations. For example, Zea mays, which is an allogamy crop, was confirmed to have short attenuation distances of < 100 kb (Gore et al. 2009; Dinesh et al. 2016). The average distance of LD decay in upland cotton was more than 3 cM (~ 1.5 Mb) in the previous studies using SSR markers (Abdurakhmonov et al. 2008; Fang et al. 2013; Li et al. 2016a). The results limit the depth exploration of QTL mapping and gene discovery, such as map-based cloning. However, the increase in the recombination rate through artificial crossing could efficiently break the genetic linkage and speed up LD decay. In our study, the distance of the LD decay of the MAGIC population was 0.76 cM (r2 = 0.1), which was estimated with 284 SSR markers. It is a relatively low LD level that represents suitable precision for QTL mapping.

Association mapping based on a MAGIC population is a powerful and reliable tool for detecting variation in genomic regions. There were three statistical models, including GLM (naive, Q, and PCA), MLM (K, Q + K, and PCA + K), and AD test, that were used in association analyses (Yang et al. 2014; Sun et al. 2016; Xu et al. 2016). In general, the population structure is the major factor that increases the number of false positives in marker–trait association (Yu et al. 2006). The convergent cross could eliminate or attenuate the effect of population structure. Given the noninterference from structure for a MAGIC population and the preponderance of MLM (Mackay et al. 2014; Sallam and Martsch 2015), we chose the MLM (K) model for the current study. Although the low density of SSR markers resulting from low polymorphism is inconvenient for delicate loci and gene mining, 52 significant loci were identified for six fiber quality traits (Table 2; Fig. 3), which provides preliminary and systematic studies for the genetic basis of fiber quality traits. In addition, MLM (K) is a reliable method for complex trait association analysis, and most of the loci were simultaneously identified by phenotypic data after BLUP as well as traits in a single environment (Table 3). The marker–trait loci identified by MLM (K) were also frequently shared with the GLM model. Furthermore, compared to the previous studies of QTL mapping in cotton using SSR markers published in an online database (http://www2.cottonqtldb.org:8081), 14 markers coincided with the previous results. Nine of them, including BNL1231, BNL3089, BNL3452, DPL0354, Gh277, HAU1129, HAU2026, NAU3419, and NAU6468, were reported to be located in QTLs for fiber traits. DPL0345 and NAU3419 were stable loci associated with fiber length that were also detected by two previous researchers (Cao et al. 2014; Jamshed et al. 2016). The SSR markers BNL1231 and BNL3452 were two hotspots mapped with multi-traits, including fiber quality, yield, and resistance traits (Table S11) (Said et al. 2015). In addition, MON_CGR5525, Gh185, BNL1231, HAU1321, MON_SHIN0376a, TMB1838, MON_DC20017, MON_DPL0441, and NBRI_HQ526817 were the new loci that control more than two fiber traits (Table S10). These loci express pleiotropism and create multiple effects in fiber development.

Along with the development of sequencing technology and genomics, association mapping has been used for candidate gene mining (Ma et al. 2016; Andres et al. 2017). The identified significant SSR markers have been anchored on physical locations that provided candidate regions in confidence interval based on LD decay distance. The genes located in these regions could be further verified by fine mapping or using reverse genetics methods. The candidate genes could also be screened by genome annotation (Sun et al. 2016). For example, the annotation gene for Gh_A06G0913 in Arabidopsis thaliana, AT5G25170, has been reported to code PPPDE putative thiol peptidase family protein which is related to cell development (Kim et al. 2013), while Gh_A06G0913 is located in MON_DPL0375c-FUHML locus (Table S12). In addition, the genes with known functions can also provide reference for candidate genes. GhPIP1-2, GhPIP2-4, GhACT1, and GhCER6, located in candidate regions of HAU4552b (SFC), MON_CGR5707 (MV), MON_CGR5352b (FE), HAU1058 (FUHML), and MON_DPL0542 (FUHML), were involved in fiber elongation or development proved by gene expression or transgenic technology (Luo et al. 2003; Li et al. 2005, 2013a; Qin et al. 2007; Liu et al. 2008). Besides, numbers of genes near the candidate markers but out of confidence regions should also be taken attention as linkage or cluster, such as GhXTH, GhFBP7, GhGA20ox2, GhKNL1, GhHOX3-A, GhPRF1_D, GhWLIM1a, GhSusA1, GhF3H, GhHOX3-D, and GhCaM7 were also important during fiber growth (Lee et al. 2010; Xiao et al. 2010; Zhang et al. 2011; Argiriou et al. 2012; Jiang et al. 2012; Han et al. 2013; Tan et al. 2013; Gong et al. 2014; Shan et al. 2014; Tang et al. 2014). The candidate loci and genes found in our study are a valuable resource for cotton genetic analyses as well as cotton improvement.

In conclusion, a cotton MAGIC population containing 960 lines was developed from 8 parents, and the population showed abundant phenotypic and genetic variation as well as a low LD level. The association mapping based on 284 SSR markers provides useful information for understanding the genetic basis of fiber traits and for genetic improvement of cotton fiber quality.