Cotton (Gossypium spp.), primarily the tetraploid species, Gossypium hirsutum L., is produced in many diverse tropical, subtropical and warm temperate regions of the world. Because cotton cultivation is so extensive, ca. 31 million hectares worldwide in 2016 (USDA-NASS), the societal benefits from plant breeding-based gains in cotton performance are greatly magnified. These include improved natural resource conservation and preservation, reduced reliance on undesirable chemical protectants, more time- and energy-efficient production, and increased profits for growers and cotton-related businesses, as well as local and even national economies. Improvements in baseline crop production and product quality have been possible through breeding of cotton germplasm resources, especially genetically elite cultivars. The use of non-elite germplasm resources for genetic improvement has been increasing in part because molecular markers and sequence polymorphisms facilitate germplasm analysis, classification, categorization, genotyping, genomic comparisons, and various types of marker-assisted selection for breeding and breeding-related research.

The importance of cotton fiber as a commodity is widely recognized due to its extensive use in textiles. The importance of cottonseed as a source of cooking oil and feed for cattle is less recognized and has been far less studied. Renewed interest in cottonseed for extended use as feed for non-ruminants or even human consumption comes from the increasing need for affordable sources of protein to feed a growing global population [1] and the ability to selectively eliminate gossypol from the seed [2]. Increased demand for feed and food will likely change across time and increase seed usage. Regardless, global climate changes will require cultivars adapted to higher temperatures, decreased precipitation, and/or increased salinity along with new genotypes with resistance to altered profiles of pests and pathogens. To meet these challenges, plant breeders will need to identify novel sources of variation and incorporate them into their breeding programs.

Cotton germplasm resources available worldwide [3] contain beneficial allelic variations that traditional and genomic breeding methods can exploit to develop cultivars adapted to emerging environmental and climate conditions. The USDA National Cotton Germplasm Collection (NCGC) is one of the largest collections of cotton germplasm resources [3]. The NCGC is comprised of over 10,000 accessions representing nine genomes and 45 species originating worldwide. However, the unequivocal differentiation, categorization, and classification of accessions remain challenging for Gossypium germplasm resources [4]. In some cases, race and species designations remain ambiguous [5–9]. Continual characterization and evaluation using the latest technologies are essential to providing the most accurate information to potential germplasm users. The ‘Gossypium Diversity Reference Set’ (GDRS) is a subset of approximately 20% of the entire NCGC, created in 2013 via proportional representation across taxonomic levels and geographic origins [4]. Simple sequence repeat (SSR) markers were recently used to genotype the GDRS to evaluate the applicability of a core set of markers across multiple genera and species [4] and to identify differences between individual accessions and groups of wild and improved types of G. hirsutum and G. barbadense L. [10]. Based on this information, core sets of G. hirsutum and G. barbadense are being developed for future genetic studies. The GDRS has been evaluated for seed oil, protein, and seed index traits [11], and subsets are being evaluated for other agronomic traits.

Diverse resources within other worldwide cotton germplasm collections have also been characterized with molecular markers, primarily SSRs [7, 12–14]. Additionally, elite cultivars from breeding programs are often genotyped to enable molecular comparisons of cultivars, known pedigrees, and the respective breeding germplasm resources in specific countries and growing regions [15–20]. The accumulation of knowledge regarding genetic differences is especially useful when combined with knowledge of differences in phenotypes because it allows breeders and other researchers to generate segregating populations of increased usefulness by choosing parents with complementary phenotypic traits and genetic distinctiveness. These populations are then utilized to combine and select favorable alleles; map and determine the genetic basis of a particular phenotype; and launch development of advanced cultivars.

With decreasing resources available for germplasm preservation, optimization of these resources becomes essential. Unrecognized redundancy of germplasm within collections has played a large role in impeding optimal management and use of collections [21]. It has been estimated that only one-third of the total number of accessions in rice germplasm banks may be distinct [22], and this situation to greater or lesser extent is recognized in many germplasm collections. Management of any plant collection requires a multitude of resources to maintain a highly viable seed inventory to ensure seed is available for utilization from the collection at any time. For cotton, the photoperiod sensitivity of many lines requires costly field and nursery production in two or more latitudes. Thus, maintenance of duplicated materials leads to considerable amounts of unnecessary effort and cost. Another concern in the management of germplasm collections and breeding programs is the maintenance of purity. Purity of lines is compromised by outcrossing, much of which cannot be easily detected by phenotypic observations. Accurate genotypic profiles in addition to phenotypic profiles are increasingly relied upon in purity testing of commercial cultivars [23]. Genotypic profiling is a sought after tool for germplasm resource management and is relevant to germplasm collections of cotton, which is self-pollinated yet highly amenable to insect-mediated cross-pollination [24].

Most recently in cotton, SSRs have been the marker of choice for researchers to identify genotypic differences [10, 12, 13, 15, 17, 25]. SSR markers typically can detect many alleles per locus, resulting in a high polymorphism which allows for utilization of a smaller number of markers. However, genotyping with SSRs is time-consuming, moderately costly, and difficult to do in a high-throughput manner [26]. In contrast, single nucleotide polymorphisms (SNPs), only have two alleles at each locus, and therefore a single marker is less informative at a locus than an SSR. However, SNPs occur more frequently than SSRs throughout the genome, and greater numbers of markers can be obtained. The benefit of SNPs is that it is possible to genotype a large number of SNPs simultaneously in a high-throughput, cost effective manner due to their bi-allelism. When diversity analyses were conducted with an equal number of SSRs and SNPs in maize, the SSR analysis was generally more informative [26]. But when the number of SNPs was increased, this limitation was overcome, and SNPs were able to resolve differences between extremely similar individuals as well as increase the accuracy of diversity estimates [26, 27].

Compared to many field crops, cotton has only recently begun to benefit from the wide availability of SNP markers. Allotetraploid crops like cultivated cotton are especially challenging for the identification of SNPs [28, 29]. Most efforts on cotton SNPs to date have focused mainly on discovery and validation [30–37]. These efforts have culminated in the identification of sufficient SNP markers to establish high-throughput genotyping systems, such as the CottonSNP63K array. These systems enable researchers to efficiently analyze large numbers of SNP markers and obtain highly repeatable results [38]. In the current investigation, we aim to utilize this SNP array for analyzing diversity within a predominantly G. hirsutum cultivar set and to identify loci associated with seed nutritional traits.

To obtain a more thorough characterization of the genetic diversity of cotton, we used the CottonSNP63K array to analyze a total of 395 cotton samples provided by the NCGC and collaborators worldwide. The data collected on the samples shows how this technology can be applied 1) to evaluate genetic diversity and population structure between groups of germplasm, 2) to gauge the effectiveness of the array to identify differences among individual G. hirsutum genotypes, 3) to investigate possible redundancies in NCGC accessions, 4) to relate this new technology with results from SSR markers, and 5) to explore the potential of the array for use in genome-wide association studies.

The development of next generation sequencing technologies and the resulting detection of thousands of markers, primarily SNPs, have increased the power available to analyze diversity in germplasm collections [21, 55, 56], and to use this diversity to make advances in plant breeding programs [57–61]. We have surveyed variation across Gossypium species utilizing the recently developed CottonSNP63K array and a large panel of 390 Gossypium samples comprising 292 improved cultivars of G. hirsutum and small sets of non-improved G. hirsutum and other Gossypium species. This data allowed us to appraise the utility of the array for evaluating genetic diversity and population structure, and to conduct SNP-based characterization and analyses on a significant portion of the USDA National Cotton Germplasm Collection. The observed patterns of diversity among Gossypium species and within G. hirsutum agreed with expectations based on previous studies with molecular markers [4, 10]. The markers distinctly separated G. hirsutum from other species and distinguished the wild from improved types of G. hirsutum. Within improved types of G. hirsutum, the observed clustering may be a sign of slight ascertainment bias as has been noted in other SNP studies [62, 63] with samples similar to TM-1 being localized towards the bottom of the two clusters of improved G. hirsutum. We found the array to be effective for grouping germplasm using PCoA and STRUCTURE analyses.

We identified a subset of 192 improved and wild cotton accessions that were run in a previous SSR analysis and on the CottonSNP63K array, and a comparison of grouping and discrimination revealed generally similar patterns for the SSR- versus SNP-based distributions. Moreover, combining the SSR data with the SNP dataset resulted in plots without marked changes, indicating that either technology is sufficient for determining relationships between samples. It is nonetheless noteworthy that the CottonSNP63K array and cluster file were primarily designed to discriminate among cultivated types of G. hirsutum and between cultivated G. hirsutum types versus non-cultivated germplasm, including several diploid and tetraploid species [38]. The ability to discriminate among non-cultivated types was not an emphasized criterion in selecting which SNPs to place in the array, but the array seems to function reasonably well for that purpose, too. The findings here indicate those objectives were achieved, in that the array and associated cluster file efficiently discerns differences among G. hirsutum cultivars, and between improved genotypes from wild forms of G. hirsutum and other species. Thus, the SNP array presents a facile means of identifying several thousand or more SNPs between most pairwise combinations of G. hirsutum genotypes. Given that the intraspecific linkage maps average about 3,500 cM [38], the expected density of markers for a given pair of parents would typically exceed one SNP per cM.

SNPs have quickly become the preferred marker system for measuring genetic diversity in plants due to the potential to identify a substantial number of SNPs throughout the plant genome and therefore have greater power when associating that information with phenotypic traits of interest. Only recently have SNPs become utilized for genotyping by the cotton research community. To be successful, SNPs need to be dense and polymorphic within the population of interest. The SNPs on the array displayed polymorphism across the range of samples tested. Over 60% of the SNPs in this dataset were found to be polymorphic in the economically important G. hirsutum species. This suggests that this array will likely have high utility for discrimination and association studies within the primarily cultivated cotton species. Specifically, discrimination by the array is best for US and Australian cultivars as they are the source of the SNPs used in development of the array. In the present study, we have shown that the SNP genotype information obtained using CottonSNP63K array has enabled GWAS to identify loci associated with a seed trait. In our analysis of cottonseed nutritional traits, four SNPs were found significantly associated with protein content, with two located in Chromosome 2, a member of the A_t subgenome. In a recent association mapping study of seed oil and protein in cotton, 228 SSR markers were used to genotype 180 elite G. hirsutum cultivars [64]. The SSR-based GWAS detected 12 markers across 9 chromosomes associated with seed protein. Interestingly, chromosome 2 that was identified with multiple individual markers in our SNP-based study was not significant in the SSR study. This may be due to much fewer SSRs tested and the corresponding low marker density relative to the number of effective SNP loci, particularly on chromosome 2. Of the eight SSR loci examined on chromosome 2 by Liu et al. [64], none co-localized with our SNPs based on alignment positions to currently available cotton genome sequences.

While multiple markers with significant associations were identified with seed protein content, no significant associations were found for seed oil content or seed index. Detection of significant marker associations is dependent upon statistical power to detect linkage disequilibrium of the marker with the causative allele for the phenotype. Statistical power in GWAS can be influenced by many factors including trait architecture and effect size, number of samples in the population, number of markers, distribution of markers, statistical method utilized, etc. [65, 66]. It may be that the effect sizes of these traits are extremely small implying a multi-genic effect which combined with a relatively small number of samples for a standard GWAS analysis would cause the inability to detect significant associations with the seed index and oil content phenotypes. The differences in trait architecture of the phenotypes is observed in the coefficient of variations (CV) and showed that the trait architecture of seed index with CV of 20.9% was very different from that of oil content (8.8%) and protein content (9.1%). This large amount of phenotypic variability in the seed index trait of the sampled population could have possibly caused QTL determination of the loci responsible for the trait to be difficult. The variability in the current study likely resulted because the data were not from a single experiment in which all of the 195 accessions were included. Rather, the data were obtained from seed harvested from a number of different seed increases grown in different locations and years [11], as in common in large germplasm collections. A quantitative trait defined in a range of environments is likely to produce different accession values that are confounded by these non-genetic effects, and it is these accession means that were used in the GWAS experiment. Inheritance studies have shown that protein and oil are both influenced by the environment, with protein being more susceptible to these changes [64, 67, 68]. Detection of significant associations with GWAS studies has generally been difficult unless a large number of samples are utilized for lowly heritable traits, as those traits that are determined by a large number of very small effect loci are very difficult to identify. Nonetheless, the utilization of standardized genotyping methods, such as possible with the array, allow for easily adding samples to the population and reanalyzing with the additional data to assist in generating additional power for detection in the future.

In addition to the GWAS performed here, an additional analysis using the CottonSNP63K array, followed by a “targeted” approach of a specific set of SNPs rather than genome-wide SNPs was recently published. Zhu et al. [69] identified a gene on chromosome 15 of the D_t subgenome that affects leaf shape in cotton; the findings were supported by similarity to genes regulating leaf morphology in other plant species. A separate study targeting SNPs for genetic male sterility (ms) in cotton [70] successfully linked a SNP to the recessive gene ms5 on chromosome 12 (A12) and ms6 on chromosome 26 (D12). Similarly, Ellis et al. [71] used the array to identify markers linked to a viral resistance gene for cotton bunchy top disease. This targeted SNP association analysis identified nearly as good an interval spanning the Cbd resistance locus as was obtained from screening an F_2:3 population. These results demonstrate that genotyping and GWAS can be effectively conducted with SNPs in cotton utilizing the array genotyping platform. As with many quantitative traits, statistical power necessary for identification of all loci responsible for a trait needs to be high to detect the small effect of numerous loci. In instances where a large proportion of the overall phenotype is explained by large numbers of loci, genomic selection utilizing the complete genotypes obtained by the array may provide a promising future avenue of research. An additional option would be to increase power by adding additional samples, as has been done with human and other crop analyses. For example, extensive application of an analogous array to the USDA Soybean Germplasm Collection enabled synthesis of haplotype block maps that are expected to lead to the discovery of many SNP-trait associations for economically important characteristics [72]. The consistency and reliability of genotyping with the array is likely to prove valuable in regard to accumulating high-quality genotypic data sets to analyze these kinds of genetic traits, with data potentially being accumulated across multiple labs, large distances and time periods.

We assessed the utility of the CottonSNP63K array data for distinguishing among G. hirsutum cultivars by comparing representative cultivars from the NCGC and worldwide cotton researchers. Because breeders often use a limited range of material, assessing the relatedness of cultivars can assist with selecting more distantly related lines to maximize variability in breeding programs. Recent SNP-based reports show a lower diversity in cotton relative to other plant species. A diversity study in soybeans (Glycine spp.) showed Chinese landraces of G. max L. (H_E = 0.338) with a slightly higher diversity than inbreds (H_E = 0.313) [73]. Though still showing higher diversity than cotton, an Italian grape (Vitis spp.) germplasm collection had higher diversity in the primary cultivated species of V. vinifera L. ssp. sativa (H_E = 0.345) than the wild form, V. vinifera L. ssp. sylvestris (H_E = 0.266) [74]. Our current evaluation with SNPs and previous evaluations with SSRs [10, 25] have revealed differences between individual cultivars with limited differences among groups of cultivars from the United States or from other countries. To maximize genetic diversity estimated with the SNP data presented here, a breeder has multiple options. A breeder could choose parents from those analyzed within this manuscript based on the genome-wide genetic similarity values calculated with all markers (Additional file 5) or could choose two cultivars on opposite ends of either the x- or y-axis shown in the MDS plots (Fig. 4a and b; Additional file 2). If a breeder was interested in genotypes not assayed here but could assign them to a breeding region as defined here, the average expected polymorphisms for a given cross could be estimated based on the category averages. For example, within the US, if germplasm from the eastern region were crossed to germplasm from the plains region, a breeder could expect to observe ca. 5,200 polymorphic SNPs between the two samples (Table 2). Breeders interested in utilizing parents developed within the US, could benefit by using cultivars from different regions in a breeding program, as they are generally less similar than cultivars from within the same region. Of particular interest are the US cultivars that could not be assigned to a specific breeding region based on regional location of breeding programs or pedigree (also includes cultivars with mixed origins). Increased heterozygosity/gene diversity and unique alleles were observed among cultivars in this group compared to cultivars with clear origin because this group was comprised of many unique genotypes. The knowledge provided by the screening of germplasm materials will allow for intelligent design of breeding populations to incorporate desired levels of known diversity particularly for targeting or pyramiding regions of interest that may provide multiple modes of action for desirable traits in the future. With continuing advances in SNP analysis technologies, the diversity data regarding specific pairs or groups of parents could also be used in concert with genetic maps to identify and select ad hoc subsets of markers to deploy for analysis and/or selection of specific families or populations. Many breeders thus might consider running the CottonSNP63K array against their breeding lines to assess their germplasm relative to other germplasm of interest, and select marker subset(s) that will prospectively maintain effectiveness but reduce genotyping costs.

While the current study is one of the first looking primarily at US germplasm, SNPs have been tested to determine their suitability in the examination of distinctness, uniformity and stability (DUS) for cotton varieties in China [75]. DUS examination typically involves a grow-out test with measurement of a set of phenotypic observations; however, the more rapid results provided by molecular markers have been allowed under the International Union for the Protection of New Varieties of Plants (UPOV) system of plant variety protection in specific situations [76]. Twenty-three core SNP markers were identified by Kuang et al. [75] that clustered 30 standard cotton cultivars into groups consistent with known pedigrees. A similar effort was undertaken to verify marker-based distinctness for registering alfalfa cultivars [77]. They compared 2,902 SNP markers with 11 morpho-physiological measures and 41 polymorphic SSRs on a set of 11 alfalfa landrace cultivars. Even though the authors identified inconsistencies for SSR and SNP markers with morpho-physiological traits, they assert that marker-based assays have high potential to substitute for, or complement, morphological distinctness measures in alfalfa cultivars. Many of these are the same issues that arise when officially registering cultivars of any crop and are applicable to identifying accessions in germplasm collections.

The integration of genomic data into germplasm collection documentation systems and its combination with taxonomic and phenotypic data will impact how these genetic resources are conserved as well as how they are used by plant breeders. In the current study, eight pairs of genotypes with high IBS values (putative duplicates based on genotype) were grown in the field and phenotypic descriptors were measured. Results of genotype and phenotype comparisons were mixed. In some cases, cultivars with a highly similar DNA profile for this set of SNP markers also had identical phenotypes, while in other cases, cultivars with a highly similar DNA profile had multiple phenotypic differences. The observed phenotypic differences are too strong to be due to natural environmental effects. It must be noted that unusual observations may be due to the possibility of human errors in any assessments, whether they be based on genotype or phenotype. Therefore, a researcher desiring to use a similar approach to remove highly similar individuals from their lines may wish to verify a small number of markers on independent DNA isolations with simplex SNP marker assays in the lab to ensure that a mistake was not made in sampling prior to more extensive field-based assessment. Based on our data at this time, we assert that any potential duplicates or redundancies within a collection cannot be ascertained solely by genotype or phenotype profiles, but that SNP and/or SSR data can bring greater power for identification of redundancy, solely due to the number of “markers or traits” one can look at. Often with phenotypic characterization, one is only working with 50–75 traits at most. SNP markers may be more appropriate to distinguish among closely related individuals while SSR markers can easily estimate diversity among less related germplasm. When comparing SSR and SNP markers in cotton cultivars, the study of Kuang et al. [75] found that SSR markers tend to correlate cultivars with their geographic origin while SNP markers more consistently correlate cultivars based on kinship or pedigree; however, no rationale was provided to explain this difference. Transcriptome profiles of lines could also potentially be different without causing an easily detectable phenotype in the field and the causal allele in most cases may not be fully represented on the marker profiles generated with the array, so a researcher may also wish to study the transcriptomic profiles of lines in the future prior to concluding lines are redundant. Until more thorough studies are available, a complement of genetic and phenotypic data should be analyzed prior to declaring two accessions as genetically distinct or duplicates.

With increasing SNP discovery projects and the development of the CottonSNP63K array to assay thousands of SNPs, it is anticipated that SNP markers will play an increasingly important role in cotton genetics, germplasm conservation, and breeding applications. In this study, we provide a large genome-wide variation data set for primarily cultivated cotton. Thousands of SNPs in representative cotton genotypes provide an opportunity to finely discriminate among cultivated cotton from around the world. The SNPs will be useful as dense markers of genome variation for association mapping approaches aimed at correlating molecular polymorphisms with variation in phenotypic traits, as well as for molecular breeding approaches in cotton.