Sequencing‐Based Bin Map Construction of a Tomato Mapping Population, Facilitating High‐Resolution Quantitative Trait Loci Detection

Abstract

Genotyping‐by‐sequencing (GBS) was employed to construct a highly saturated genetic linkage map of a tomato (Solanum lycopersicum L.) recombinant inbred line (RIL) population, derived from a cross between cultivar NC EBR‐1 and the wild tomato S. pimpinellifolium L. accession LA2093. A pipeline was developed to convert single nucleotide polymorphism (SNP) data into genomic bins, which could be used for fine mapping of quantitative trait loci (QTL) and identification of candidate genes. The pipeline, implemented in a python script named SNPbinner, adopts a hidden Markov model approach for calculation of recombination breakpoints followed by genomic bins construction. The total length of the newly developed high‐resolution genetic map was 1.2‐fold larger than previously estimated based on restriction fragment length polymorphism (RFLP) and polymerase chain reaction (PCR)–based markers. The map was used to verify and refine QTL previously identified for two fruit quality traits in the RIL population, fruit weight (FW) and fruit lycopene content (LYC). Two well‐described FW QTL (fw2.2 and fw3.2) were localized precisely at their known underlying causative genes, and the QTL intervals were decreased by two‐ to tenfold. A major QTL for LYC content (Lyc12.1) was verified at high resolution and its underlying causative gene was determined to be ζ‐carotene isomerase (SlZISO). The RIL population, the high resolution genetic map, and the easy‐to‐use genotyping pipeline, SNPbinner, are made publicly available.

Abbreviations

BIL

backcross inbred line

FW

fruit weight

GBS

genotyping‐by‐sequencing

HMM

hidden Markov model

IM

interval mapping

LOD

logarithm of odds

LYC

fruit lycopene content

NIL

near‐isogenic line

PVE

phenotypic variation explained

QTL

quantitative trait loci

RIL

recombinant inbred line

SGN

Sol Genomics Network

SNP

single nucleotide polymorphism

VCF

variant call format

Core Ideas

• Construction of a GBS‐based high‐resolution genetic map of a tomato RIL population

• Development of an analytical pipeline to localize crossover events and construct genomic bins

• Fine mapping of QTL for fruit weight and lycopene content confirm use of the high‐resolution map

• Identification of SIZISO as the gene underlying the lycopene QTL Lyc12.1

• RIL population, high‐resolution genetic map, and analytical pipeline available to the public

Recombinant inbred line populations are powerful tools for plant genetic studies. Recombinant inbred lines are usually derived from a cross between two inbred parents followed by a series of self‐pollinations and generation advancement until complete or nearly complete homozygosity is achieved (Page and Grossniklaus, 2002). Each resulting RIL displays a unique genetic structure that is a mosaic of the two parental lines in both genotypic and phenotypic characteristics. The permanent, immortal nature of a RIL population provides ample opportunities for the population to be phenotypically evaluated under different conditions for various traits contrasting in the two parental lines. A linkage map of a RIL population allows genetic investigation of the contrasting traits under different conditions. For example, RIL populations are very suitable for the identification and characterization of QTL for traits segregating in the population (Broman, 2005). As multiple genomic introgressions are present in homozygous conditions in the RILs, such populations enable the detection of additive genetic effects as well as epistatic (additive × additive) interaction effects of contributing loci. In tomato, several RIL populations were previously developed, mostly by crossing a cultivated tomato line with either wild tomato species (Villalta et al., 2005; Foolad, 2007; Capel et al., 2015; Foolad and Zhang, 2015; Demirci et al., 2017) or S. lycopersicum var. cerasiforme (Alef.) Voss accessions (Saliba‐Colombani et al., 2000; Alba et al., 2009; Pascual et al., 2015). These RIL populations were mostly genotyped by classic molecular markers such as RFLPs and simple sequence repeats.

With recent advancements in sequencing technologies, sequencing‐based markers, such as SNPs, are becoming more common for genetic mapping and QTL analysis as they provide more detailed map information. For example, using whole genome resequencing data, Huang et al. (2009) demonstrated that ultra‐high‐resolution genotyping of a rice (Oryza sativa L.) RIL population allowed QTL detection with 100 kbp resolution. Comparison of a traditional RFLP map with a resequencing‐based map in rice demonstrated that the latter map not only detected more QTL per trait, it also identified QTL for grain length and grain width that were an order of magnitude narrower than the intervals identified using the RFLP‐based map (Yu et al., 2011). Whole‐genome resequencing was also employed to resolve the genomic location of a root‐knot nematode resistance QTL to a 27 kbp interval (harboring two candidate genes) in a soybean [Glycine max (L.) Merr.] RIL population using a bin‐based mapping approach (Xu et al., 2013). These studies were feasible mainly because of reductions in sequencing costs. A more recently developed and still evolving technique is GBS first introduced by Elshire et al. (2011). This technique uses a methylation‐sensitive restriction enzyme for DNA fragmentation to generate reduced representation libraries of the genome followed by highly multiplexed sequencing of mapping populations. Genotyping‐by‐sequencing has complemented SNP arrays as an efficient tool for genetic studies in an increasing number of plant species including maize (Zea mays L.), barley (Hordeum vulgare L.), wheat (Triticum aestivum L.), rice, sorghum [Sorghum bicolor (L.) Moench], soybean, and cassava (Manihot esculenta Crantz) (Poland and Rife, 2012; Spindel et al., 2013; Glaubitz et al. (2014) and references therein; Rowan et al., 2015). Genotyping‐by‐sequencing has been used widely in biparental mapping populations such as F₂ generations, RILs, and backcross inbred lines (BILs) as well as in natural populations. In tomato, GBS was employed in a BIL population derived from a cross between tomato breeding line M82 and S. pennellii Correll accession LA0716 (Fulop et al., 2016). In these studies, the reduced representation libraries of the genome covered by the sequencing reads produced sufficient numbers of markers to generate high‐density maps.

While GBS and resequencing data can provide hundreds of thousands of SNPs for biparental populations, a medium‐size RIL population will probably contain only a few thousands genomic bins (e.g., Edwards et al., 2016). False SNP calling, a result of varying amounts of read depth per sample and the possibility of large amounts of missing data, may make the calculation of homologous recombination points and construction of genomic bins a challenging task. To overcome such challenges, researchers have tried various methods of sliding window scripts using both SNP windows (Huang et al., 2009; Zhao et al., 2010; Zhou et al., 2016; Demirci et al., 2017) and physical windows (Lin et al., 2014; Han et al., 2016). While these methods are based on sliding window principles, they vary in window sizes and step lengths, with each species inherently having a unique and different genome size and structure. Using a different approach, Xie et al. (2010) employed a hidden Markov model (HMM) algorithm to create a bin‐based map of a rice RIL population, concluding that an average of 8× coverage should be used to avoid genotyping errors. Hidden Markov model was also employed to create a bin‐based genetic map for a soybean RIL population using resequencing data with 0.19× effective read depth, achieving a mean of 20.8 kbp recombination interval (Xu et al., 2013). Rowan et al. (2015) used HMM methodology with GBS data to detect crossover events in an Arabidopsis thaliana (L.) Heynh. F₂ mapping population and mapped a QTL for flowering time to a 9 kbp interval harboring its causative gene cluster, MAF2‐5. Hidden Markov model was also used in a tomato BIL population, facilitating narrowing of QTL intervals in several locations (Fulop et al., 2016).

In a previous study, we reported the development of a RIL population of 170 lines, derived from a cross between tomato breeding line NC EBR‐1 and S. pimpinellifolium accession LA2093, and a corresponding genetic linkage map based on 294 RFLP and simple sequence repeats markers (Ashrafi et al., 2009). In a subsequent study, we used the population and its genetic map to identify QTL for various fruit quality characteristics, including FW, pH, total soluble solids, and LYC (Ashrafi et al., 2012). Seven QTL for FW were detected on six tomato chromosomes, with confidence intervals of 14 to 33 cM, which were in general agreement with FW QTL reported previously in various other tomato mapping populations (reviewed by Grandillo et al., 1999). Major LYC QTL identified in our previous study included Lyc7.1 on chromosome 7 and Lyc12.1 on chromosome 12, with the latter showing a much larger effect on LYC content and localized within a confidence interval of ∼6 cM (Ashrafi et al., 2012). In a further study, using marker‐assisted backcrossing and near‐isogenic lines (NILs), the Lyc12.1 QTL interval was narrowed to a region of ∼1.5 cM, representing a length of 1.5 Mbp, and a candidate gene approach suggested a SlZISO gene as a possible candidate (Kinkade and Foolad, 2013).

In the present study, we employed a GBS approach to develop a high‐resolution genetic map of the aforementioned RIL population, hypothesizing that an increase in the number of genetic markers would allow fine mapping of QTL for many traits segregating in the RIL population. Thus, a set of 141,083 polymorphic SNPs were identified differentiating the genomes of NC EBR‐1 and LA2093. Using HMM, we calculated crossover events and generated a bin‐based genetic map comprising 2869 genomic bins. The high‐resolution genetic map allowed not only significant reductions in interval sizes for several previously identified QTL, but also identification of most probable candidate genes underlying major QTL. The RIL population as well as the high‐resolution genetic map and SNP markers are available at the C.M. Rick Tomato Genetics Resource Center (http://tgrc.ucdavis.edu/) and Sol Genomics Network (SGN; https://solgenomics.net/), respectively. Moreover, the SNPbinner package for calculating recombination breakpoints and associated scripts are available at the SGN GitHub page (https://github.com/solgenomics/snpbinner).

MATERIALS AND METHODS

Plant Material

A previously described RIL population of tomato, developed at Penn State University from a cross between S. lycopersicum breeding line NC EBR‐1 and S. pimpinellifolium accession LA2093 (Ashrafi et al., 2009), was used for GBS analysis and genetic bin mapping. Because of the scope of the present study, only 148 lines of the RIL population were used. The 148 RILs at F₈ generation were grown in a greenhouse at the University of California Davis, and leaf samples were collected from individual RILs and used for DNA extraction and GBS analysis (described below). Seven high‐LYC NILs, previously developed at Penn State University in the NC EBR‐1 background (Kinkade and Foolad, 2013), and their parental lines NC EBR‐1 and LA2093, were grown under greenhouse conditions (with a 16‐h light period) at Cornell University, Ithaca, NY. At least three plants of each NIL (except NIL‐544 with two plants) were grown for fruit analysis. Ripe fruit at 7‐d post breaker stage were harvested and pericarp tissue from three fruits per plant was flash frozen in liquid nitrogen, pooled, and stored at −80°C for subsequent carotenoids and RNA analyses (described below).

Genotyping‐by‐Sequencing Analysis and Single Nucleotide Polymorphism Calling in Parental and Recombinant Inbred Lines

Genomic DNA was isolated from leaf tissues of the 148 RILs and two parental lines using DNeasy MiniPrep extraction kit (Qiagen, https://www.qiagen.com/us/) and digested with restriction enzyme NlaIII. Reduced representation libraries were constructed according to Monson‐Miller et al. (2012) and multiplexed and sequenced on three lanes of an Illumina HiSeq 2500 platform for paired‐end 2 × 100 bp reads. Reads were demultiplexed into individual sample files, using 5‐bp inline barcodes and a script from the Comai lab at UC Davis (available from http://comailab.genomecenter.ucdavis.edu/index.php/Barcoded_data_preparation_tools), and subsequently processed for quality control and trimming (using CLC Genomics Workbench v. 6.0.5). Trimmed reads of the parental lines and 148 RILs were aligned to the tomato genome assembly v. 2.50 (Tomato Genome Consortium, 2012; Fernandez‐Pozo et al., 2015; https://solgenomics.net/) using CLC Genomics workbench with default parameters (except for the following: ignore multi‐hit mapping, read length fraction 0.6, identity fraction 0.9). Alignment results were exported in the form of BAM files for all samples. The BAM files for the parental lines were used for SNP calling through SAMtools v. 0.1.7a and custom Perl scripts, as outlined in Hulse‐Kemp et al. (2016). The positions of all possible SNPs derived from this pipeline were used to call genotypes for each of the RILs using the Genome Analysis Toolkit (GATK) v. 3.4–46 (McKenna et al., 2010) with HaplotypeCaller and genotyping mode GENOTYPE_GIVEN_ALLELES with output mode EMIT_ALL_SITES. The resulting variant call format (VCF) file was filtered initially to select for positions where the parental samples had different genotypes. The VCF file was subsequently further filtered using GATKs SelectVariants program for minor allele frequency >0.1 and samples having genotype information >80%. The final VCF file was imputed using Beagle 4.1 (Browning and Browning, 2016) and transformed into mapping data format with ABH using a custom Perl script available from GitHub (https://github.com/solgenomics/SNPbinner/blob/master/other_scripts/vcf_to_csv.pl).

RNA Sequencing

For analysis of expression levels in the parental lines and high‐LYC NILs, ripe fruit pericarps were used. RNA extraction, library construction, and sequencing were performed according to Zhong et al. (2011), except that the Illumina HiSeq2500 instrument was used for sequencing. Analysis of RNA‐sequencing reads was performed according to Rosli et al. (2013), except that reads were aligned to v2.50 of the tomato genome and raw counts were obtained using ITAG2.4. The edgeR (Robinson et al., 2010) functions glmQLFit and glmQLFTest (Lun et al., 2016) were combined with the makeContrasts function to examine differentially expressed genes between the parental lines as well as contrasts for each NIL to LA2093 and to NC EBR‐1.

Breakpoints Calculation and Population Genotyping using SNPbinner

Transitions between genotypes (breakpoints) along the chromosomes in each RIL, as well as genotype bins, were calculated using the SNPbinner application developed in the present study. A full description of SNPbinner is available as Supplemental File S1. Briefly, analysis was performed in three parts, each implemented as a subcommand in a python script, as illustrated in Fig. 1 and described below.

Fig. 1.

Workflow of SNPbinner pipeline.

The first subcommand, crosspoints, uses a pair of HMMs, one accounts for heterozygosity while the other does not, to determine the most likely genotype at each point along a chromosome for each RIL, given user‐defined emission and transition probabilities. Here we used an emission probability value of 0.9 (an error probability of 0.1), representing the prediction of genotypic homogeneity of each region. Then, continuous genotype regions, which fall below a user‐defined minimum size (0.2% of chromosome length used in this study), are subsumed by surrounding regions using a simple greedy algorithm (described in Supplemental File S1). The breakpoints and genotype of each region for each RIL are then reported as the output of the crosspoints subcommand.

The second subcommand, visualize, is used to create image representations of the original SNPs against the calculated breakpoints in each chromosome in each line (see Supplemental File S1). That step is used for quality control and to determine the optimal parameters of crosspoints to be used. The user can define any combination of input and output files to facilitate manual verification of the results.

The third subcommand, bins, uses the aggregate breakpoints resulting from crosspoints to create representative bins for the whole population and determines the genotype for each RIL within each bin. This subcommand uses a modified k‐means clustering algorithm (described in Supplemental File S1) to group neighboring breakpoints while also maintaining a user‐specified minimum distance (5 kbp used in this study) between the centers of each group. The mean of these aggregate breakpoint groups are then used as boundaries to create a map of bins. Each RIL is then projected onto this map to determine the genotype of each bin of a given RIL. The projections of each RIL are then reported as the output of the bins subcommand to be used in further analysis.

Quantitative Trait Loci Analysis

Our previous phenotypic data for FW and LYC content, collected from four generations of the RIL population (Ashrafi et al., 2012), were used in the current study for QTL analysis. Data were first examined for normality of distribution and transformed to log₁₀ or square‐root to reach normality, if necessary. For QTL analysis, R/qtl (Broman et al., 2003) was employed with an automated script published by Spindel et al. (2013). Briefly, interval mapping (IM) was performed using the scanone function, then, additional linked QTL were searched (addqtl) followed by forward and backward selection and elimination model using stepwiseqtl function. This process in R/qtl is a version of composite interval mapping, which can reduce the background noise of QTL detected by IM particularly when multiple QTL are present on the same chromosome. Subsequently, the model that could best explain the phenotypic variation (fitqtl) was adopted. For each trait, the significance level was calculated using 1000 permutations and α = 0.05. A few steps were added to the script to include several genotyping explorations that plotted the recombination frequencies and displayed the logarithm of odds (LOD) scores from the model into a file. To be consistent with our previous QTL analyses (Ashrafi et al., 2012), 2‐LOD support intervals were calculated for all QTL. The upgraded R/qtl script is made available along with SNPbinner on GitHub (https://github.com/solgenomics/snpbinner).

Carotenoids Extraction

Carotenoids were extracted from 200‐mg finely ground frozen pericarp tissue according to Alba et al. (2005). Subsequently, the extract was dried at reduced pressure and resuspended in ethyl acetate containing 50 μg ml⁻¹ of diindolylmethane as an internal standard. Carotenoids were separated and quantified using an ultra‐performance convergence chromatography system, equipped with a HSS C18 SB column (2.1 by 150 mm, 1.8 μm particles size, Waters, http://www.waters.com). The column effluent was monitored between 250 and 500 nm wavelength using a photodiode array detector. The flow rate was 1 mL × min⁻¹ starting from 99% eluent A (supercritical CO₂) and 1% eluent B (Methanol) with a linear gradient over 7.5 min to 20% B, which was held for 4.5 min followed by linear gradient back to 1% B at 15 min. Individual carotenoids were resolved in comparison with authentic standards (Sigma). Beta‐carotene was used to construct a five‐point calibration curve (10–140 ng μL⁻¹) and the relative concentrations of each carotenoid are reported in β‐carotene equivalents.

Data Availability

The raw reads of the GBS from 148 RILs and parental lines are available through NCBI short reads archive project number PRJNA449767. The raw reads of the RNA sequencing from the seven high‐LYC NILs and LA2093 and NC EBR‐1 are available through NCBI short reads archive project number SRP139770. The genetic map and genotyping table are available via SGN.

RESULTS

Genotyping‐by‐Sequencing, Binning, and High‐Resolution Genetic Map Construction

The GBS analysis of the RIL population (n = 148 lines) and its two parental lines resulted in a genome coverage of 9.2× for NC EBR‐1, 2.7× for LA2093, and an average of 2.3× for the RILs (Supplemental Fig. S1). A total of 144,520 polymorphic SNPs were detected, of which 141,083 were physically mapped onto the 12 tomato chromosomes and 3437 to the artificial chromosome 00 harboring sequences not yet localized to one of the tomato chromosomes. For binning, we used only the SNPs mapped to the 12 tomato chromosomes. Genomic bins were generated using the SNPbinner application, which implemented HMM in the subcommand crosspoints to calculate recombination breakpoints. The performance of crosspoints was visually evaluated against the original SNP dataset using visualize subcommand of SNPbinner. Based on visual evaluations, the calculated breakpoints were generally in good agreement with the SNP data (Fig. 2; for complete evaluation see Supplemental Dataset S1). Because of the excess residual heterozygosity in 18 RILs (>10% of the genome, expected <1.5% in F₈), they were eliminated and the remaining 130 RILs were used in subsequent analyses. The average and median number of breakpoints per chromosome in each RIL was determined to be 2.6 and 2, respectively (Table 1; Supplemental Table S1). The number of total recombination events per RIL varied from 13 to 59, with an average of 30.8 and a median of 29, and the total number of recombination events in the whole population (130 RILs) was 4007 (Supplemental Table S1). In all chromosomes, the recombination events occurred predominantly in genic regions, showing association with high gene density (Fig. 3).

Fig. 2.

Visual evaluation of recombination breakpoint calculation by SNPbinner. The position of transition from LA2093 genotype (red) to NC EBR‐1 genotype (blue) and vice versa, as calculated by SNPbinner, is plotted in comparison to the original SNPs data in chromosome 3 of selected individual RILs. The upper panel depicts the original SNPs and the lower panel depicts the calculated breakpoints. The full evaluation of the calculated breakpoints can be viewed in Supplemental Dataset S1. Blue SNPs or regions are NC EBR‐1, red are LA2093, green are heterozygous and white are no data.

Table 1.

Recombination events occurred in the entire population (130 recombinant inbred lines [RILs]).

	Per RIL per chromosome	Per RIL
Average	2.6	30.8
Median	2	29
Third quartile	4	36
Minimum	0	13
Maximum	12	59

Fig. 3.

Recombination breakpoints frequency vs. gene density. The frequencies of the recombination breakpoints (blue dots) in their genomic location in each one of the 12 chromosomes as calculated by SNPbinner are presented in respect to the gene density in each position (yellow bars, derived from SGN website). Breakpoints locations were rounded to the nearest 100 kbp.

A bin‐based high‐resolution genetic map was constructed using SNPbinner subcommand bins. The resulting map was comprised of 2869 bins with an average bin size of 280 kbp and a median of 57 kbp. Approximately 75% of the bins were shorter than 111 kbp, and each bin harbored on average 12.2 genes with a median of six genes. The recombination frequencies were plotted with MadMapper (http://cgpdb.ucdavis.edu/XLinkage/MadMapper/), which revealed absence of genetic linkage between different linkage groups or between nonadjacent bins (Supplemental Fig. S2). The only exception was for the pericentromeric region of chromosome 12, where it appeared that an inversion might have occurred between the two species or that this area in the genome was assembled in reverse orientation. When the bins in this region were positioned in reverse orientation, the recombination frequencies were aligned without any obvious anomaly (Supplemental Fig. S3). This region of chromosome 12 accounted for the entire scaffold no. 5 of the chromosome, and its reversed orientation was validated by in situ fluorescence hybridization (Shearer et al., 2014; Saha et al., 2016).

The genetic map distances were calculated with R/qtl using Kosambi mapping function, and the total map length was estimated to be 1362 cM (Table 2; Supplemental Fig. S4). This is 259 cM longer than the RFLP‐based map (1103 cM), constructed with the same subset of 130 RILs. The average map distance between adjacent bins was 0.48 cM with a median of 0.39 cM compared with 3.53 and 2.4 cM, respectively, in the RFLP‐based map (Ashrafi et al., 2009). This represents an average of sevenfold decrease in distance between markers. The genotyping table is presented as Supplemental File 2.

Table 2.

Genetic distances (cM) calculated for the population using sequencing data and restriction fragment length polymorphism (RFLP) markers. Distances were calculated with R/qtl using Kosambi mapping function.

Chromosome	GBS bin‐based map	RFPL‐based map	Fold change
01	160.1	127.7	1.25
02	125.0	93.1	1.34
03	138.5	96.3	1.44
04	107.0	97.7	1.10
05	111.2	93.6	1.19
06	71.2	69.7	1.02
07	103.8	95.9	1.08
08	103.8	86.5	1.20
09	122.8	86.8	1.42
10	103.5	78.0	1.33
11	111.9	101.5	1.10
12	103.4	76.3	1.35
Entire map	1362.4	1103.2	1.23
Avg. distance between markers	0.48	3.91	0.12

The genetic composition of the entire population included 46.3% cultivated parent genome, 52.1% wild parent, and 1.6% heterozygous regions (Supplemental File S2; Supplemental Fig. S5). Skewed segregations from the expected Mendelian ratio were observed in a few genomic regions including a large region (46 cM, 58 Mbp) on chromosome 3, where only ∼22% of the RILs displayed the NC EBR‐1 genotype (expected 50%), and a few smaller regions on chromosomes 5, 6, and 10. On chromosomes 5 and 6, skewed segregations were observed in favor of NC EBR‐1 in a 35‐cM region (5.4 Mbp) and a 39‐cM region (10 Mbp), respectively. On chromosome 10, there was a 4‐cM region (0.8 Mbp) displaying skewed segregation in favor of the wild‐type parent LA2093. Similar deviations from the expected 50% ratio were reported in the RFLP map of this RIL population (Ashrafi et al., 2009).

Quantitative Trait Loci Analysis using the High‐Resolution Map

We examined the utility of the new genetic map for QTL analysis for FW and LYC content, two traits for which QTL were previously identified in four generations (F₇, F₈, F₉, and F₁₀) of this RIL population using an RFLP‐based genetic map (Ashrafi et al., 2012). The complete data of the QTL mapping generated by the R/qtl script is available as Supplemental Dataset S2 and briefly described below.

Fruit Weight

The average values for FW across the four generations (4 yr) were 2.2 g for LA2093 and 160.7 g for NC EBR‐1 with the RIL population FW ranging from 3.9 to 52 g with an average of 12.1 g (Ashrafi et al., 2012). The best QTL models for FW detected, respectively, six, five, six, and five QTL in the F₇, F₈, F₉, and F₁₀ generations. The percentage of total phenotypic variation explained (PVE) by the models was 61 to 77% (Table 3). In all generations, the most prominent QTL for FW was detected on chromosome 2 at location 101 to 110 cM with a confidence interval of 90 kbp to 1.7 Mbp (Table 3, Q4). This QTL interval harbors 11 to 254 genes, with 18 to 25% PVE across the four generations. The detected QTL intervals in F₇, F₉, and F₁₀ generations all harbored Solyc02g090730, the causative gene underlying QTL fw2.2 determined in a previous study (Frary et al., 2000). Moreover, in F₇ and F₁₀ generations the peak of the QTL was in the bin harboring Solyc02g090730 (ch02_52191227..52260920) along with 10 other genes. The interval in F₈ generation ends 300 kbp away (accounting for one bin) from the gene Solyc02g090730.

Table 3.

Fruit weight quantitative trait loci (QTL) calculated in four consecutive generations (F₇–F₁₀). All QTL were generated with multiple QTL models of R/qtl.

QTL	Q1	Q2	Q3	Q4	Q5	Q6	Q7	Q8	Q9	Q10	Q11	Q12	Model
Chromosome	1	1	1	2	3	4	4	5	9	11	11	12	–
Position (cM)†	14.4	33	45.4–48.9	101.4–110.1	98.7	45.8	65	18.6	97.6	56.9	94.0–97.1	19.3–23.6	–
Interval size (cM)‡	17	6.1	20.6–53.2	1.5–12.4	2–6.8	10.3	14.4	22.5	30.2	11.3	4.7–19.9	16.9–21.2	–
Interval size (Mbp)‡	1.5	1.3	3.7–78.8	0.09–1.7	0.17–1.1	48.2	2.1	1.5	3.3	2.9	0.31–1.7	1.0–1.7	–
LOD/PVE§, F7	5.2/5.1	–	–	19.1/24.9	14.5/17.2	–	–	4.5/4.3	4.1/3.9	–	12.5/14.1	–	38.2/76.7
LOD/PVE, F8	–	8.5/10.2	–	14.2/19.2	10.3/12.9	–	–	–	–	–	13.3/17.7	5.2/5.9	34.6/70.9
LOD/PVE, F9	–	–	7.1/9.6	13.0/19.7	7.9/10.8	5.7/7.5	–	–	–	6.2/8.2	–	4.4/5.6	30.8/69.1
LOD/PVE, F10	4.3/7.1¶			9.9/18.1	9.6/17.4	–	4.6/7.6	–	–	–	5.8/9.7	–	24.5/61.3

^†Position of marker with highest logarithm of odds (LOD) score.

^‡Two‐LOD support interval size.

^§Percentage of variance explained.

^¶Interval size span over all three QTL and its intervals exceeds the intervals of individuals QTL.

A QTL on chromosome 3 was detected in all generations, positioned precisely at 98.7 cM, displaying 10 to 17% PVE for FW (Table 3, Q5). The confidence interval for this QTL varied from 167 kbp to 1.1 Mbp including 20 to 150 genes. In all generations, the QTL interval harbored SlKLUH (Solyc03g114940), which encodes a cytochrome P450 enzyme accounting for QTL fw3.2 determined in previous studies (Zhang et al., 2012; Chakrabarti et al., 2013). Furthermore, for this QTL, the interval peak in all four generations was at bin ch03_64688578..64779804, which harbors SlKLUH along with 10 other genes.

While the model‐selection approach detected one QTL on chromosome 11 in all four generations (Table 3), the IM approach (scanone function) identified two to three QTL in different generations (Fig. 4A; Supplemental Table S2). The peak of the first QTL, detected only by IM, was located at 49.6 cM, encompassing 2.2 Mbp with 13 to 19% PVE (calculated based on a single QTL model). This QTL was colocalized with QTL fw11.1, previously identified (Grandillo et al., 1999). The second QTL was located at 95.5 cM, encompassing 1.6 Mbp and PVE of 12 to 19%. This QTL was detected in the multiple‐QTL‐models (F₇, F₈, and F₁₀), at 94 to 97.1 cM encompassing an average interval length of 0.3 to 1.7 Mbp and PVE of 10 to 18% (Table 3, Q11). The third QTL on chromosome 11, detected only by IM approach and mainly in F₇ and F₈ generations (a peak below threshold was detected for this QTL in F₉ and F₁₀ generations; Fig. 4A), was located at position 106.8 cM, encompassing an interval of 8.6 cM (1 Mbp) and PVE of 16%. This region coincides with the fascinated gene (Solyc11g071810; Cong et al., 2008) and the fw11.3 locus harboring the cell size regulator gene, SlCSR (Solyc11g07194; Huang and van der Knaap, 2011; Mu et al., 2017). The peak for the QTL was at the bin ch11_55192430..55352524, harboring SlCSR along with 24 other genes.

Fig. 4.

Fruit weight quantitative trait loci (QTL) analyses using interval mapping (IM) approach for four generations in selected chromosomes. (A) Chromosome 11, (B) Chromosome 1. The QTL scan was performed with R/qtl using scanone using the EM algorithm. Horizontal lines represent significant cutoff calculated with 1000 permutations and α = 0.05. Data was log₁₀ transformed for normalization.

Interval mapping detected at least two FW QTL on chromosome 1 in all four generations (Fig. 4B; Supplemental Table S2), whereas multiple QTL models detected only one (Table 3, Q1–3). The first QTL peak was detected at 10 cM, spanning 0.5 to 30 cM in different generations (∼3.8 Mbp) and PVE of 13 to 20%. This QTL was detected by the multiple‐QTL model only in F₇ generation (Table 3, Q1). The second QTL on this chromosome was located at 52 cM, spanning 45 to 65 cM and 3.2 Mbp region and PVE of 13 to 17%. This QTL was detected by the multiple‐QTL model only in the F₉ and F₁₀ generations (Table 3, Q3).

A QTL was detected on chromosome 12 by multiple‐QTL models, significantly in F₈ and F₉ generations and just below significant threshold in F₇ and F₁₀ generations (Supplemental Table S2). This QTL positioned at ∼20 cM with ∼5.5% PVE (Table 3, Q12) and spanning 1.7 Mbp. In addition to the above major QTL, several small‐effect QTL (PVE ≤ 7.6%) were detected by multiple‐QTL models in one or two generations on chromosomes 4, 5, and 7 (Table 3).

Lycopene

Based on the colorimetric model, LA2093 accumulated 110.1 μg g⁻¹ FW LYC, whereas NC EBR‐1 accumulated 28.2 μg g⁻¹ FW (Ashrafi et al., 2012). The average LYC values across generations of the RILs were 32.1 to 130.0 μg g⁻¹ FW with a median of 68.9 μg g⁻¹ FW. In the present study, we confirmed all the LYC QTL identified in our previous study, including major QTL Lyc12.1 and Lyc7.1 (Ashrafi et al., 2012). However, the high‐resolution QTL analysis reduced the confidence interval of Lyc12.1 to as small as 2 cM (Table 4) and its physical interval size to 1.5 to 1.8 Mbp in different generations. The 1.5 Mbp interval comprised seven bins (249 genes), with the bin harboring the prime candidate gene, SlZISO (encoding ZISO enzyme), was the largest bin, encompassing 431 kbp and 70 genes. The phenotypic variation explained by Lyc12.1 ranged from 19.8 to 23.2%. The QTL Lyc7.1 was detected in F₇ and F₈ generations with interval size of 1.1 to 1.4 Mbp, explaining 14% of the total phenotypic variation for LYC content. Additional smaller QTL were detected on chromosomes 1 and 10 in F₁₀ generation. The total LYC phenotypic variation explained by the multiple‐QTL models in the different generations varied from 34 to 43% (Table 4).

Table 4.

Lycopene quantitative trait loci (QTL) calculated in four consecutive generations (F₇–F₁₀). All QTL were generated with multiple QTL models of R/qtl, lycopene levels measured with colorimetric assay.

QTL	Q1	Q2	Q3	Q4	Model
Chromosome	1	7	10	12
Position (cM)†	46.1	4.7–5.1	18.5	114.3–115.5
Interval size (cM)‡	34.3	8.8–47	13.3	2–13.2
Interval size (Mbp)‡	76.1	1.0–58.1	1.25	1.5–2.7
LOD/PVE§, F7	–	5.1/14.1	–	6.9/19.9	10.9/34.0
LOD/PVE, F8	–	5.5/13.6	–	8.7/22.8	13.2/37.5
LOD/PVE, F9	–	–	–	5.0/20.5	–
LOD/PVE, F10	3.6/8.8	–	4.2/10.2	9.7/26.6	13.7/41.3

^†Position of marker with highest logarithm of odds (LOD) score.

^‡Two‐LOD support interval size.

^§Percentage of variance explained.

Validation of the Causative Gene under Lyc12.1 using Near‐Isogenic Lines

To further validate the hypothesis that SlZISO was the causative gene underlying Lyc12.1 (Kinkade and Foolad, 2013) and to understand SlZISO function in LYC biosynthesis in the RIL population, we analyzed the carotenoid levels and the transcriptomes of seven high‐LYC NILs and the two parental lines of the RIL population. In these experiments, LA2093 accumulated 2.5‐fold higher LYC than NC EBR‐1 (87 and 35 μg g⁻¹ FW, respectively), and the NILs accumulated 1.5‐ to 2.5‐fold more LYC than NC EBR‐1. For five NILs, this difference was statistically significant (Table 5), though all NILs accumulated more LYC than NC EBR‐1 (Supplemental Table S3). Differential gene expression analysis, focusing on known carotenoid structural and regulatory genes, indicated that 11 of the genes were differentially expressed between LA2093 and NC EBR‐1 (Supplemental Table S4, S5). Among them, the only gene for which the NILs showed expression similar to LA2093 (high expression) and different from NC EBR‐1 (low expression) was SlZISO (Fig. 5; Supplemental Table S5). Specifically, the log₂ fold change of SlZISO between NC EBR‐1 and LA2093 was 6.19, and among NC EBR‐1 and the NILs ranged from 5.95 to 6.48. This clearly indicated the involvement of SlZISO in production of high LYC concentrations in LA2093 and the high‐LYC NILs. Further analyses were conducted to determine the potential causes of low expression of SlZISO in NC EBR‐1. These included an examination of SNPs on chromosome 12, which were common to NC EBR‐1 and several wild tomato accessions for which sequence data were available. Using only the polymorphic SNPs of the studied population, we discovered that the last ∼2 Mbp of chromosome 12 of NC EBR‐1 displayed SNPs in common with S. habrochaites accession LA1777 (Fig. 6). This finding indicates potential introgression of this region from S. habrochaites accession PI 126445, one of the parental lines used to develop NC EBR‐1 (Gardner, 1988).

Table 5.

Carotenoids levels (μg × g FW^{−1 †}) in LA2093, NC EBR‐1, and high‐LYC near‐isogenic lines (NILs). All values are mean of three biological replicates (NIL‐544, n = 2) ± SE.

	Lycopene	β‐carotene	Phytoene	ζ‐carotene	Lutein
LA2093	87.4 ± 10.5*	8.1 ± 0.4*	4.1 ± 0.5*	1.2 ± 0.3*	1 ± 0.1*
NC EBR‐1	35.3 ± 5.1	2.4 ± 0.2	9.4 ± 3.4	3.9 ± 0.2	0.5 ± 0.03
NIL‐1139	88 ± 5.5*	3.3 ± 0.3*	3.6 ± 0.8*	0.8 ± 0.05*	0.6 ± 0.04
NIL‐1425	84.9 ± 17*	2.1 ± 0.1	4.7 ± 0.6	0.8 ± 0.1*	0.4 ± 0.2
NIL‐1432	83.3 ± 19.9*	2 ± 0.3	3.9 ± 0.2*	0.6 ± 0.04*	0.5 ± 0.04
NIL‐1460	80.2 ± 1.6*	2.4 ± 0.1	5.1 ± 0.3	0.6 ± 0.3*	0.5 ± 0.04
NIL‐140	65.9 ± 11*	4 ± 0.3*	5.1 ± 0.9	0.8 ± 0.1*	0.5 ± 0.04
NIL‐1438	52.6 ± 5.1	2.2 ± 0.1	4.2 ± 0.3*	0.7 ± 0.1*	0.4 ± 0.04
NIL‐544	54.8 ± 9.5	1.8 ± 0.5	3.2 ± 0.3*	0.7 ± 0.1*	0.5 ± 0.1

^†FW, fresh weight; all carotenoids were calculated as β‐carotene equivalents

^*values significantly different from NC EBR‐1 (α < 0.05, student t‐test).

Fig. 5.

Relative expression of carotenoid biosynthesis genes in ripe fruits (breaker + 7 d) in the population parents and in the high‐LYC near‐isogenic lines (NILs) as determined by RNA‐seq. Different letters indicate lines for which expression is significantly different in comparisons with LA2093, as determined by edgeR analysis. Values are RPKM (reads per kilobase of exon model per million mapped reads) and represent the mean ± SE of at least three biological replicates (two for NIL‐544). CCD1B, carotenoid cleavage dioxygenase 1B; ZDS, ζ‐carotene desaturase; PSY, phytoene synthase; ZISO, ζ‐carotene isomerase.

Fig. 6.

Analysis of chromosome 12 SNPs in NC EBR‐1 parent. (A) Count of SNPs common to NC EBR‐1 and S. habrochaites LA1777 (+), S. pennellii LA0716 (×), S. chilense CGN15532 (♦), and S. galapagense LA483 (●). (B) Zoom‐in on the last 7.2 Mbp. All SNPs are relative to the tomato reference genome. The SNPs used for the analyses are the population polymorphic SNPs and the public data from Sol Genomics Network published by Aflitos et al. (2014). Mbp, mega base‐pairs.

We also retrieved the coding sequences of SlZISO from both parental lines based on the RNA sequencing data and compared them with SlZISO from the tomato reference genome (cv. Heinz 1706, v2.5; www.solgenomics.net). While LA2093 contained only one SNP compared with the reference genome (common also to NC EBR‐1), NC EBR‐1 displayed 10 SNPs with two of them in heterozygous state (Supplemental Fig. S6). On the amino acids level, the common SNP resulted in a V104I substitution in comparison to Heinz, and NC EBR‐1 showed additional four amino acid substitutions in comparison to both LA2093 and Heinz (Supplemental Fig. S7).

Analyses of other carotenoids indicated that, the level of β‐carotene was almost four times higher in LA2093 than in NC EBR‐1, whereas in the NILs β‐carotene level was similar to that in NC EBR‐1 (Table 5). This observation confirms the report that the NILs were specifically developed for high fruit LYC content (Kinkade and Foolad 2013), further suggesting presence of a different regulation mechanism for β‐carotene accumulation than the total flux downstream of the SlZISO enzyme. Compounds upstream of LYC in the carotenoids biosynthesis pathway showed a more complex picture. For example, ζ‐carotene accumulated in the NILs at a level similar to LA2093, while phytoene, the first compound in LYC biosynthesis pathway, did not show any difference between LA2093 and NC EBR‐1 and was accumulated in significantly lower levels in the high‐LYC NILs (Table 5).

DISCUSSION

Sequencing‐based genotyping has been widely employed in various populations and plant species because of increasing output and reduced cost (Xie et al., 2010; Elshire et al., 2011; Gao et al., 2013; Spindel et al., 2013; Xu et al., 2013; Glaubitz et al., 2014; Hill et al., 2015; Li et al., 2015; Han et al., 2016). To use these reduced representation sequencing data in biparental populations, we developed SNPbinner application, which calculates recombination breakpoints using HMMs and generates an amalgamated genotyping table of the entire population. The HMM‐based binning tools that are currently available online provide pipelines that comprised of comprehensive data analyses with multiple steps including SNP calling (Xie et al., 2010; Fulop et al., 2016). SNPbinner, however, is designed specifically for the post‐SNP detection part of the analysis to simplify the binning process and to allow the users to choose the SNP calling platform that best fits their study from the various publicly available tools (Yi et al., 2014). SNPbinner was also designed to use manual HMM parameterization rather than a training set to allow intuitive adjustment of its performance. SNPbinner's crosspoints function uses the last contiguous‐genotyped SNP locus as a RIL breakpoint rather than using imputation to estimate the breakpoint between the two boundary SNPs as in previous software (Xie et al., 2010). This feature provides direct rather than imputed results, allowing SNPbinner's bins function to create more precise means between breakpoints across the RIL population. While this can reduce breakpoint distance compensation to sites of interest between the boundary SNPs of genotype regions, the high density of the SNPs in our analysis minimizes the error that this might cause. In the current study we did not set heterozygous regions to missing data as was previously done (Xie et al., 2010; Li et al., 2015), which might give a more comprehensive estimation of the allelic status of each RIL across the bins. In addition, SNPbinner contains a visualized quality control step that we found very helpful to determine the optimal values for the user‐defined parameters. Altogether, these make SNPbinner an intuitive, flexible, and easy to optimize tool.

Using SNPbinner, we developed a GBS‐based high‐resolution genetic map of a tomato RIL population derived from a cross between a cultivated tomato and an accession of S. pimpinellifolium, the most closely related wild species of tomato and the only red‐fruit tomato wild species. The genetic map includes 141,083 SNPs in 2869 genomic bins anchored on the 12 tomato chromosomes. The high‐density genetic map spanned over 1362 cM, which is larger than the estimated 1103 cM based on ∼300 RFLP and PCR‐based markers (Ashrafi et al., 2009). This is most likely a result of detecting additional crossover events, leading to increased map resolution and refinement. Observations of map expansion when sequencing data are used, compared with the use of low‐density traditional markers, have been reported in other plant species (Wang et al., 2010; Yu et al., 2011). In tomato, the well‐studied introgression line population (n = 75 lines) derived from a cross between tomato cultivar M82 and S. pennellii accession LA716, dissected the tomato genome into 107 genomic bins when using RFLP markers (Eshed and Zamir, 1995; Liu and Zamir, 1999). Using the same parental lines, Ofner et al. (2016) developed a BIL population in the background of cultivar M82 consisting of 446 lines, and by using SNP‐chip‐based genotyping found 633 bins. Further, when the number of BILs was expanded to 545 lines, and genotyped by GBS analysis, 1049 genomic bins were detected, greatly increasing the map resolution. We have observed a similar phenomenon in the present study, but since RILs are much more variable than BILs and each line harbors multiple introgressions, we have been able to dissect the tomato genome into many more genomic bins by using only 130 lines. These bins will allow precise mapping of QTL for any trait contrasting between LA2093 and NC EBR‐1.

Using the same RIL population including all 170 lines, Ashrafi et al. (2012) mapped the two largest‐effect QTL for FW on chromosomes 2 and 3 to intervals of 21 and 34 cM, respectively. Here, by employing the high‐density map and using only a subset of 130 RILs, the QTL interval on chromosome 2 was reduced to ≤12.4 cM in all four generations (1.5 cM in F₇), harboring the known causative gene for fw2.2 (Frary et al., 2000). Similarly, the major FW QTL on chromosome 3 was mapped to an interval of 2 to 7 cM in the four generations, and the peak of the interval included the bin harboring the known fw3.2 locus and its underlying causative gene, SlKLUH (Zhang et al., 2012; Chakrabarti et al., 2013). These observations demonstrate the power of GBS‐based high‐resolution genetic map for finer mapping of QTL and more precise localization of candidate genes. The GBS‐based IM approach also allowed detection of several smaller‐effect QTL for FW, including those on chromosomes 1 and 11, which were previously reported in one or more tomato populations (Chen et al., 1999; Grandillo et al., 1999; Ashrafi et al., 2009). In particular, the QTL covering fw11.3, for which the causative gene was recently reported to be SlCSR (Mu et al., 2017), resides in the bin on chromosome 11 with the highest LOD score in all four generations.

Ashrafi et al. (2012) also identified a high‐LYC QTL on chromosome 12 (Lyc12.1) within ∼6‐cM interval. In a subsequent study, using high‐LYC NILs in the background of NC EBR‐1, Lyc12.1 was fine‐mapped to a region of about 1.5 cM and a candidate gene approach suggested potential involvement of SlZISO as the causative gene underlying this QTL (Kinkade and Foolad, 2013). Here, we mapped Lyc12.1 to an interval of ∼2 cM, comprising seven bins encompassing ∼1.5 Mbp and 249 genes. Fantini et al. (2013) had demonstrated the involvement of SlZISO in LYC accumulation in tomato fruit, using a virus‐induced gene‐silencing approach. To examine the potential role of SlZISO in high fruit LYC accumulation in LA2093 and its high‐LYC NILs, we employed a differential gene expression approach. The expression analysis, focusing on major known carotenoid structural and regulatory genes indicated that SlZISO was expressed in LA2093 and its derived high‐LYC NILs ∼80 times higher than its expression in the low LYC parent NC EBR‐1. No other carotenoid gene was found to be so differentially expressed among these lines. These results strongly support the suggestion that SlZISO is most likely a major cause of high LYC production in LA2093 and its derived high‐LYC NILs, though whether it is the only contributing factor could not be determined in this study. While it is possible that the AA changes observed in SlZISO from NC EBR‐1 compared with LA2093 influence its activity, which seems negligible compared with its differential expression, which is possibly a result of additional sequence variations in the promoter region. The ZISO enzyme catalyzes the isomerization of 9‐15‐9′‐tri‐cis‐ζ‐carotene to 9–9′‐di‐cis‐ζ‐carotene (Chen et al., 2010), which is the isomer required by the subsequent enzyme, carotenoids isomerase (CRTISO), for further conversion toward LYC (Isaacson et al., 2002). This conversion, however, occurs spontaneously in the presence of light (Chen et al., 2010), which may explain the greater redness of the peripheral pericarp in most tomato cultivars. The greater expression of SlZISO in LA2093 and its derived high‐LYC NILs is most likely responsible for the higher LYC concentration in the interior fruit tissues of these lines. Based on the combined mapping and differential expression analysis results we propose SlZISO as the gene underlying LYC QTL Lyc12.1.

CONCLUSIONS

Combining reduced‐representation low‐depth sequencing with HMM‐based binning process, facilitated the development of a high‐density, bin‐based, genetic map of a tomato RIL population. The use of this genetic map for QTL analysis in the present study resulted in precise identification of QTL regions and determination of their underlying candidate genes for important fruit quality traits. The strategy developed and used in the current study could be employed in a cost‐effective manner in other populations of both self‐ and cross‐pollinated crops to associate phenotypes of interest with underlying candidate genes. The availability of the germplasm with its genetic data as well as the software developed, will be beneficial for other researchers.

Supplemental Information Available

Supplemental information is available with the online version of this manuscript.

CONFLICT OF INTEREST

The authors declare that they have no conflict of interest.

DISCLAIMER

Mention of trade names or commercial products in this publication is solely for the purpose of providing specific information and does not imply recommendation or endorsement by the USDA. The USDA is an equal opportunity provider and employer.

Supporting information

Supplemental Information

Supplemental File 1

Supplemental Information

Supplemental File 2

Supplemental Information

Supplemental File 3

Supplemental Information

Supplemental File 4

Supplemental Information

Supplemental File 5

Supplemental Information

Supplemental File 6

Supplemental Information

Supplemental File 7

2018‐12‐13. The Plant Genome 1 12 December 13, 2018