Introduction

Common bean (Phaseolus vulgaris L.) is an important legume crop for human nutrition, being an important source of protein, complex carbohydrates, fiber, and beneficial minerals for millions of individuals worldwide (Broughton et al. 2003; Gepts et al. 2008). The species belongs to a large and diverse genus that comprises 70–80 species, five of which have been domesticated (Freytag and Debouck 2002). Among these domesticated species, common bean is the one with the broadest geographic distribution and the highest agronomic, nutritional, and economic value (Gepts 2014). It is a diploid species with a haploid complement of 11 chromosomes and a genome size of ~587 Mb (Schmutz et al. 2014).

Repeated experimental evidence highlights the existence of two different and genetically divergent wild gene pools in common bean, called Mesoamerican and Andean gene pools, which underwent domestication independently (Bitocchi et al. 2013; Gepts 1998; Kwak and Gepts 2009; Schmutz et al. 2014) and diversified into distinct eco-geographic races (Singh et al. 1991; Chacón et al. 2007). Indeed, the Andean gene pool is generally adapted to relatively higher altitudes and lower temperature, while the Mesoamerican gene pool is adapted to lower altitudes and higher temperatures (Beebe et al. 2011). A range of molecular markers have been developed and employed in beans for the analysis of genetic diversity (domestication, gene pool divergence, and population structure), linkage mapping and association studies, and marker-assisted selection (MAS) in breeding programs (Blair et al. 2009; Kwak and Gepts 2009; Miklas et al. 2006; Talukder et al. 2010). However, marker development and use remain relatively expensive and the coverage of available markers in the genome is still modest (Varshney et al. 2014).

Next-generation sequencing (NGS) technologies are revolutionizing genetic studies and molecular markers development by exponentially increasing the number of genetic variants that can be discovered in a single experiment (Stapley et al. 2010). With these technologies, single nucleotide polymorphism (SNP) and insertion–deletion (InDel) detection and genotyping have become feasible on a whole-genome scale and are widely applied to diversity and association studies in plants (Thudi et al. 2012; Varshney et al. 2014). Nevertheless, in spite of the reduced cost of sequencing technologies and the increased throughput and multiplexing, the cost of sequencing and genotyping large numbers of individuals is still prohibitive in plants with complex and repetitive genomes (Davey et al. 2011; Descham and Campbell 2010).

Several complexity reduction approaches that couple restriction enzyme (RE) genome digestion with NGS and SNP calling have been developed in the last years for high-throughput molecular marker discovery in different organisms (Davey et al. 2011). These approaches include reduced representation libraries (RRLs) (Altshuler et al. 2000), restriction site-associated DNA sequencing (RAD-Seq) (Baird et al. 2008), restriction enzyme sequence comparative analysis (RESCAN) (Monson-Miller et al. 2012), and genotyping-by-sequencing (GBS; Elshire et al. 2011).

GBS is a robust, high-throughput, cost-effective, and simple technique for obtaining thousands of markers from large numbers of individuals. It has been applied in genetic diversity studies to both plants and animal species (De Donato et al. 2013; Elshire et al. 2011; Glaubitz et al. 2014). In addition, in spite of the high percentage of missing data (Glaubitz et al. 2014; Beissinger et al. 2013), GBS technology has demonstrated its usefulness in the identification of quantitative trait loci (QTLs) in several crops like barley, soybean, chickpea, wheat, and common bean (Hart and Griffiths 2015; Iquira et al. 2015; Li et al. 2015; Liu et al. 2014; Jaganathan et al. 2015). Despite its several advantages, GBS requires a species-specific optimization regarding the RE used to avoid repetitive regions of the genome and to determine marker number, distribution, and depth (Beissinger et al. 2013). For example, Hart and Griffiths (2015) found good SNP coverage in common bean using ApeKI, but there was uneven density distribution, probably because ApeKI is a methylation-sensitive enzyme. On the other hand, Zou et al. (2014) employed a methylation-insensitive enzyme (HaeIII) in common bean, but detected a high proportion of the SNPs (~77 %) in repetitive regions. In the research reported here, an in silico analysis of different RE was performed to identify suitable enzymes for GBS in common beans, based on the availability of a P. vulgaris reference genome sequence (Schmutz et al. 2014). We then tested the GBS method with a panel of 18 wild and domesticated P. vulgaris accessions. Results are considered in light of read mapability among genotypes, marker distribution, and sequence depth. We evaluate also the possibility of using GBS with CviAII for identifying copy number variations (CNVs) across different genotypes. The information reported here will be useful for planning other GBS experiments in common bean using a larger number of genotypes, for both diversity and association studies.

Materials and methods

In silico digestion, library preparation, and sequencing

Thanks to the availability of the P. vulgaris whole-genome sequence (Schmutz et al. 2014), a survey of different RE and their relative cutting sites could be performed. Using the Biopython suite (Cock et al. 2009), we selected enzymes that create a ‘sticky’ end after cleaving, cut only once for each recognition site, and do not recreate the restriction site after digestion. Elshire et al. (2011) suggested a methylation-sensitive enzyme to avoid repetitive elements of the genome when using GBS with maize, a plant with a large genome composed mainly of transposable elements (Schnable et al. 2009). In contrast, common bean has a relative small genome, with only 50 % of the genome belonging to pericentromeric regions, which contain 26 % of the genes (Schmutz et al. 2014). In addition, because of possible genotype-dependent differences in DNA methylation (Grativol et al. 2012), which could bias genotyping, we followed another approach. For each selected enzyme, we counted the number of recognition sites in the masked (where all the repetitive sequences are converted into string of Ns) and unmasked genome sequences, and kept those enzymes that preferentially cut in the non-repetitive part of the genome, based on a binomial test. In this subset of enzymes, we selected CviAII (recognition site C’ATG), because this enzyme showed the higher restriction site count and displayed a preferential localization in the non-repetitive part of the genome. Since ApeKI has been recently applied in common bean (Hart and Griffiths 2015), we also compared the in silico distribution of digested fragments suitable for sequencing (50–350 bp long) between ApeKI and CviAII across the genome.

In order to check the applicability of the GBS protocol using CviAII, a test experiment was performed with 17 wild and domesticated P. vulgaris genotypes belonging to both Andean and Mesoamerican gene pools. In addition, a representative of the wild ancestral gene pool from northern Peru, G21245, was also included. As internal control for our analysis, we included also the common bean genotype used for generating the genome reference sequence (G19833; Schmutz et al. 2014). Specific barcodes and adapters for CviAII were designed with the GBS barcoded adapter generator (http://www.deenabio.com/services/gbs-adapters) (Supplementary File S1).

DNA was extracted from freeze-dried bean leaves of greenhouse-grown plants using a modified protocol of Pallotta et al. (2003) with an extra step consisting in re-suspension with 4 μl of RNAse and incubation for 30 min at 37 °C. DNA quality was checked with NanoDrop Lite (Thermo Fisher Scientific) and by 1 % agarose gel electrophoresis. DNA with an absorbance ratio (A260/A280) >1.7 and with no visible degradation on agarose gel was used for subsequent library preparation. Genomic DNA and library adapters were quantified with QUBIT dsDNA HS assay kit (Thermo Fisher Scientific/Invitrogen, Grand Island, NY). GBS libraries and adapters were prepared following the protocol of Elshire et al. (2011), using CviAII (New England Biolabs, Ipswitch, MA) for DNA digestion and a 1:4 dilution of adapter mix (common and barcoded adapter) at a final concentration of 4.5 ng per reaction. In the ligation step, we reduced the ligation buffer concentration to 0.6× per reaction, instead of the suggested 1×. During the fragment enrichment step, four separate PCR amplifications were performed and the different reactions were then pooled for PCR purification. The presence of adapter dimers in the sequencing libraries was checked with the Experion DNA analysis kit (Biorad, Berkeley, CA). Genomic libraries were sequenced in a single lane of Illumina HiSeq 2000 flowcell, using the 50-bp cycle protocol, in the QB3 Vincent J. Coates Genomics Sequencing Laboratory at the University of California, Berkeley, CA. The raw sequencing reads have been deposited in the NCBI Sequence Read Archive (http://www.ncbi.nlm.nih.gov/sra) under the accession number SRX1308469.

Sequencing preprocessing, alignment, and SNP calling

Recently, TASSEL-GBS (Glaubitz et al. 2014), a specific algorithm for analysis and SNP calling of GBS datasets, was released. The software was specifically implemented for calling the maximum number of SNPs in low coverage and highly multiplexed datasets, favoring allelic redundancy over quality score (Glaubitz et al. 2014). Since our dataset contained few lines at high coverage, we preferred to follow a different, more robust, and accepted pipeline for bioinformatic analysis (Altmann et al. 2012). In particular, we used SAMtools for SNP calling since different studies indicate that it is more conservative in variant calling compared to other algorithms, also in datasets obtained from reduced representation libraries (Altmann et al. 2012; Greminger et al. 2014).

Reads were quality-trimmed at the 3′-end using sickle (https://github.com/najoshi/sickle), keeping only reads with no more than 2 Ns and a minimum length after trimming of 30 bp. Then, the reads that recreated the CviAII cutting site (possible chimeras, partial digestion, or sequencing errors) or that contained the common adapter sequence (short fragments) were trimmed and only those reads longer than 30 bp after this second trimming step were retained. The last filtering step kept only the reads that contained, after the barcode sequence, the overhang sequence of CviAII digestion (i.e., ATG). The resulting reads were then de-multiplexed using sabre (https://github.com/najoshi/sabre) allowing one mismatch for each barcode.

Read alignment was performed on the P. vulgaris unmasked genome sequence (Schmutz et al. 2014: G19833 accession; http://www.phytozome.net/commonbean) using BWA (Li and Durbin 2009). After the alignment, only the reads with a minimum mapping quality of 10 were used for downstream application. Base call recalibration was performed with the R package (www.r-project.org) ReQON (Cabanski et al. 2012). After quality score recalibration, variants were called with SAMtools considering only loci covered by more than 30 % of the lines analyzed (six lines). The resulting variants were filtered with VCFtools (Danecek et al. 2011); only those with a Minor Allele Frequency (MAF) higher than 0.05, a minimum quality more than 10, and a mean read depth, across all lines, from 5 to 1000 (–maf 0.05 –minQ 10 –min-meanDP 5 –max-meanDP 1000) were considered for downstream analysis. SNP and InDel statistics were performed with VCFtools; SNP density and transition to transversion ratio (Ts/Tv) were calculated for non-overlapping bins of 1 Mb.

Identification of repetitive regions and phylogenetic analysis

SNPs located in repeated regions were removed with VCFtools using the annotation of P. vulgaris repeats available in Phytozome (Goodstein et al. 2012).

For phylogenetic analysis, only the variants located in annotated coding DNA sequences (CDSs) were used, since these regions are generally subjected to higher evolutionary pressure than noncoding DNA sequences. A FASTA multiple alignment file was created for subsequent phylogenetic analysis by concatenating the extracted variants at each position for each genotype analyzed. During the creation of the multiple alignment file, individual genotypes with a quality below 10 or missing genotypes were treated as missing data. Due to the self-pollinating nature of P. vulgaris, the heterozygous calls were also treated as missing data, since they could be sequencing or SNP calling errors. The resulting multiple alignment file was then analyzed using the seaview toolkit (Gouy et al. 2010). A phylogenetic tree was built using the Neighbor-Joining (NJ) clustering approach, with the Kimura two-parameter (Kimura 1980) nucleotide substitution model and 1000 bootstrap replicates using the seaview toolkit (Gouy et al. 2010).

CNV identification and annotation

CNVs were identified using the reference genotype G19833 as baseline for identifying coverage shifts, as a proxy of CNV, in the other sequenced genotypes. First, we calculated the number of reads in 100-Kb non-overlapping genomic bins in each genotype. Then, we normalized the read counts in each bin by dividing the count by the total number of reads mapped in each genotype, and calculating the relative read coverage (RRC) as a ratio between the normalized read counts of the genotype of interest and the reference genotype (G19833). The RRC should be normally distributed with a mean of ~1. For this analysis, we removed the genomic bins without mapped reads in the G19833 genotype. We selected as putative CNV the genomic bins with a RRC <0.1 or >1.9; the genes located in these genomic bins were then subjected to Gene Ontology (GO) enrichment analysis using the Blast2Go tool (Conesa et al. 2005).

Results and discussion

In silico genome digestion and analysis of high-throughput sequencing raw data

Comparison of in silico genome digestion between CviAII and ApeKI showed that CviAII would produce more fragments suitable for sequencing but that it will—as expected—require a higher sequencing coverage than ApeKI. On the other hand, by using CviAII, we would be able to tag 97 % of the genes present in P. vulgaris genome, 30 % more than when using ApeKI (Table 1).

Table 1 Comparison P. vulgaris genome in silico digestion and distribution of fragments suitable for sequencing between CviAII and ApeKI
Full size table

Sequencing on a HiSeq 2000 (Illumina, San Diego, CA) generated 137,026,622 50-bp single-end reads of which 127,384,853 (93 %) passed the initial sickle quality trimming. Among these ~127 M reads, 3,002,729 (2.4 %) were removed because they were shorter than 30 bp after the trimming of reads containing the RE recognition site or adapter contaminants, or because they did not contain the overhang RE sequence after the barcode sequence. As expected from the library preparation strategy, there was a high level of duplicated reads, with only 13,278,501 unique reads in the dataset, suggesting a mean 10× redundancy for each read tag. Nevertheless, these data suggest that the overall library quality was high and consistent with the experimental approach.

After de-multiplexing, alignment, and filtering of the low-quality aligned reads, the number of reads was almost uniformly distributed across the different genotypes, with a coefficient of variation (CV) around the mean of 28 %. The majority of annotated genes in P. vulgaris genome (>90 %) were tagged by at least one read (Table 2). The ability to tag the majority of common bean genes could be useful for a more complete genotyping, given an increase in read coverage, since GBS library preparation is more cost-effective than standard NGS protocols. In addition, this characteristic could be useful when applying RE genome reduction approaches coupled with target sequence captures, like the recently developed RAD capture (Rapture) protocol (Ali et al. 2016). Almost 50 % of the reads in each line could be aligned to the reference genome; and 50 % of the aligned reads tagged gene sequences. A similar mapping efficiency was observed in other studies using GBS in common bean with different REs (Schröder et al. 2016). The total number of reads per gene in each line ranged from 36 to 84, with a mean of 52 reads per gene in each line. These results are consistent with the in silico digestion of P. vulgaris genome. Furthermore, they showed a homogeneous read mapping rate among wild and domesticated races belonging to different gene pools (Table 2).

Table 2 Distribution of de-multiplexed reads among different individuals
Full size table

Analysis of identified SNPs and InDels

A total of 77,595 SNPs and InDels was identified after keeping variants with a Minor Allele Frequency (MAF) higher than 0.05 (–maf 0.05), a minimum calling quality higher than 10 (–minQ 10) and a mean read depth per sites between 5 and 1000 (–min-meanDP 5, –max-meanDP 1000). Among the variants identified, 73,656 (95 %) were SNPs, 2088 (3 %) were deletions, and 1851 (2 %) were insertions. The InDels ranged from 1 to 8 bp, with the majority of them being mononucleotide insertions and deletions. Due to the repetitive nature of most plant genomes and the resulting miscalls of SNPs and InDels in repetitive regions, all the variants that were located in these regions were removed. The remaining number of variants were 47,838 (61 % of the total), with 23,273 variants (31 % of the total) located in genic sequences. These variants were divided between 44,875 (94 %) SNPs, 1940 (3 %) deletions, and 1693 (3 %) insertions. The ratio of non-repetitive variants is similar to the occurrence of CviAII recognition sites in non-repetitive versus repetitive regions of the genome, highlighting the reliability of in silico digestion-based approaches. In addition, the percentages of variants located in non-repetitive regions and in genic sequences were three times higher than the variants identified by Zou et al. (2014) in common bean. For further analysis, only these non-repetitive SNPs were considered.

The SNP and InDel distributions were significantly highly correlated with chromosome length (r = 0.79, p = 0.004) (Supplementary File S2), with a mean of ~4328 and a median of 4312 variants per chromosome, and a median of 79 variants per Mb. These results exceeded markedly the ones obtained after ApeKI digestion of Hart and Griffiths (2015). In particular, they found a correlation of 0.45 between SNPs density and chromosome length using the ApeKI RE in common bean. The highest number of variants were observed on chromosome 2 (5311) and the lowest on chromosome 10 (3314). On the other hand, no significant correlation was found between mean SNP density (in 1 Mb non-overlapping bins) and chromosome length (r = −0.35, p = 0.28) (Supplementary File S2). The variant mean read depth for each line ranged from 5 to 12 reads per site, with a mean and median of ~8 reads for SNPs. The variant coverage, averaged across all the lines, ranged from 5 to 439, with a mean and median of 8 and 7, respectively. A plot of variant density in 1-Mb non-overlapping bins closely resembled the density of annotated genes in the P. vulgaris chromosomes (Supplementary File S3), with a Pearson correlation coefficient (r) of 0.89 (p < 2.2e−16).

SNPs were classified into transitions (Ts) and transversions (Tv), based on the type of nucleotide substitution, using VCFtools (Table 3). The number of C/T and A/G transitions was similar (~13,000); the A/C and G/T transversions had a similar frequency, while A/T and C/G transversions were slightly higher or lower, respectively, compared to A/C and G/T transversions. The Ts/Tv ratio in our dataset was 1.56 for the SNPs localized in non-repetitive regions, slightly higher than previously reported in common beans using a RRLs approach (Zou et al. 2014).

Table 3 Transition and transversion counts for the identified SNPs
Full size table

Characterization of SNP and InDel distribution and phylogenetic analysis

The total number of SNPs and InDels per line ranged from 3512 to 21,415, with the lower number of SNPs and InDels identified in genotypes G19833 (3512), UC0801 (5354), CAL143 (5479), and Midas (9033) (Table 4). All these genotypes were domesticated beans belonging to the Andean gene pool, as does the genotype used for the Schmutz et al. (2014) reference sequence (G19833), which was also the one with the fewest SNPs in our analysis. SNPs and InDels in Mesoamerican entries ranged from 17,308 (accession PI417653) to 19,664 (PI311859 or G35101). The genotype with the highest number of variant sites was G21245, a wild bean from an ancestral gene pool originating in northern Peru (Kami et al. 1995), with 21,416 variants detected.

Table 4 SNP and InDel distributions among different genotypes and genomic features
Full size table

Of the 47,838 SNPs and InDels identified, 23,273 (49 %) were located in genic sequences, with 11,163 in CDS, 2285 in untranslated regions (UTRs), and 9825 in introns (Table 4). For all the genotypes analyzed, 45–49 % of the SNPs and InDels were located in genic sequences; among them ~50 % were located in CDS, ~40 % in introns, and ~10 % in UTRs. The 23,273 SNPs and InDels located in genic sequences identified 11,027 different genes (or 40 % of genes identified in the whole-genome reference sequence), with an average of two variants per gene.

The phylogenetic analysis based on the identified SNPs and InDels was consistent with the division in different gene pools and domesticated/wild lines, and was also significantly supported by high bootstrapping values (Fig. 1). The Andean and Mesoamerican gene pools were clearly divided with a bootstrap support >95. In particular, both domesticated groups of Andean and Mesoamerican genotypes were strongly supported by a bootstrap value of 100, confirming the major bottleneck that occurred during each of the two independent domestications of common bean (Bitocchi et al. 2013; Gepts 1998; Schmutz et al. 2014). In addition, the phylogenetic tree was automatically rooted with the ancestral genotype G21245 from northern Peru (Kami et al. 1995). Overall, the phylogenetic analysis of the variants identified using GBS with CviAII correctly identified genetic relationships among the accessions included in this study, and the level of genetic diversity of the respective gene pools based on previous information about this species (Bitocchi et al. 2013; Gepts 1998; Kwak and Gepts 2009; Schmutz et al. 2014).

Fig. 1
figure 1

Neighbor-Joining (NJ) phylogenetic tree based on variants located in genic sequences of the different bean lines. Bootstrap values and gene pools of the different lines are shown. PhI ancestral wild, DA domesticated Andean, WM wild Mesoamerican, DM domesticated Mesoamerican

Full size image

CNV identification and annotation

CviAII, having a 4-bp recognition sites, is a frequent-cutting enzyme and shows a diffuse read coverage across the genome (Supplementary File S4). Thus, this enzyme could be suitable for identifying CNVs across different genotypes with GBS and could also represent a cost-effective approach for identifying this kind of variations in different bean genotypes. Indeed, CNVs are extremely important in plant genome evolution, but also affect plant phenotypes and resistance to both biotic and abiotic stresses (Żmieńko et al. 2014). Reduced representation libraries were used previously for identifying putative CNVs in plants and animals. As example, Henry et al. (2015) used RESCAN libraries for identifying large chromosomal rearrangements and CNVs in Populus plants, while De Donato et al. (2013) identified some known CNVs using GBS in different cattle breeds. The approach used in our study showed a RRC that was normally distributed, with a mean approximately equal to 1 (Supplementary File S5), suggestive of the reliability of this approach for the identification of CNV in common bean. Analysis of RRC showed 162 genomic bins, containing 343 genes, which could contain potential CNVs in the genotypes analyzed, with some of them shared across different genotypes (Supplementary File S6). GO enrichment analysis of these genes highlight a significant enrichment in genes involved in the apoptotic process, innate immune response, transmembrane signaling receptor activity, signal transduction, ATP binding and protein binding (Supplementary File S7). A large number of these genes are annotated as leucine-rich repeat proteins and transmembrane kinases, NB-ARC domain-containing disease resistance protein, TIR-NBS-LRR class proteins, and cysteine-rich receptor-like kinases (Supplementary File S8). These observations suggest that the majority of putative CNVs segments identified in these genotypes contain genes involved in biotic stress response. This result is in agreement with previous studies in several plants that identify regions harboring CNVs as enriched in biotic stress-response genes (Cook et al. 2012; deBolt 2010; McHale et al. 2012; Żmieńko et al. 2014), further highlighting the feasibility of CNVs identification using GBS with a frequent-cutting enzyme.

Conclusions

GBS is a simple, cost-effective, and highly multiplexed protocol for plant genotyping using NGS technologies. Using this protocol, we were able to identify 47,838 variants in 18 wild and domesticated bean genotypes. Even though the use of a frequent-cutting, methylation-insensitive enzyme will require a higher genome sequencing coverage, the small genome size of common bean and the results presented in this study clearly show the advantages of using CviAII for GBS in this species. We identified thousands of evenly spaced markers across the entire common bean genome, with a high density that closely resembles genes distribution. This high density could help in narrowing QTL regions in mapping experiments and facilitating a more precise location of recombination events. In addition, 50 % of the variants identified lay in genic sequences, while the others were situated in the noncoding part of the genome. The variants in genic sequences reliably identified known phylogenetic subdivisions in common bean. They could also be useful in genome-wide association studies (GWAS) for identifying candidate genes responsible for traits of interest. On the other hand, the variants in the noncoding parts of the genome could be useful—as predominantly neutral markers—for ecological studies in this species, in particular for population modeling and for inferring demographic history in wild common bean. Our approach also allowed us to identify several putative CNVs that could be involved in pathogen response and resistance in different common bean genotypes. Last but not least, the increased throughput and reduced cost of sequencing technology will soon leverage the cost and depth of sequencing required when using GBS with different REs such as 4-bp recognizing, methylation-insensitive enzymes, especially for plants with small genomes like common bean.