Linkage Mapping and Comparative Genomics of Red Drum (Sciaenops ocellatus) Using Next-Generation Sequencing

Developments in next-generation sequencing allow genotyping of thousands of genetic markers across hundreds of individuals in a cost-effective manner. Because of this, it is now possible to rapidly produce dense genetic linkage maps for nonmodel species. Here, we report a dense genetic linkage map for red drum, a marine fish species of considerable economic importance in the southeastern United States and elsewhere. We used a prior microsatellite-based linkage map as a framework and incorporated 1794 haplotyped contigs derived from high-throughput, reduced representation DNA sequencing to produce a linkage map containing 1794 haplotyped restriction-site associated DNA (RAD) contigs, 437 anonymous microsatellites, and 44 expressed sequence-tag-linked microsatellites (EST-SSRs). A total of 274 candidate genes, identified from transcripts from a preliminary hydrocarbon exposure study, were localized to specific chromosomes, using a shared synteny approach. The linkage map will be a useful resource for red drum commercial and restoration aquaculture, and for better understanding and managing populations of red drum in the wild.

Next generation sequencing (NGS) has provided a set of powerful tools for characterizing the genome of nearly any species. However, genome assembly remains a challenge for organisms with large, repetitive genomes, such as those of many plants and vertebrates (Bradnam et al. 2013). For many applications, a suitable alternative to genome sequencing and assembly is construction of a genetic linkage map. While linkage maps traditionally have been possible only for species with pedigree information, or those that could be crossed experimentally, a combination of reduced-representation NGS (Baird et al. 2008;Elshire et al. 2011) and advances in single-cell genomics (Xu et al. 2015) have made it possible to obtain a high-density linkage map for almost any species. High-density linkage maps generated recently have been used to understand the genetic basis of adaptive phenotypic traits (Baxter et al. 2011;Henning et al. 2014), study the genetic basis of sex determination (Anderson et al. 2012;Palaiokostas et al. 2013a,b), map quantitative trait loci in economically important species (Chutimanitsakun et al. 2011;Houston et al. 2012;Talukder et al. 2014), and for population and comparative genomics Hohenlohe et al. 2012;Bradbury et al. 2013;Manousaki et al. 2015).
The red drum (Sciaenops ocellatus) is an estuarine-dependent, marine fish distributed throughout the northern Gulf of Mexico (GoM) and along the east coast of the United States (Pattillo et al. 1997). The species supports a large recreational fishery in U.S. waters and is reared in captivity for both restoration aquaculture in the U.S. and commercial aquaculture in China, the U.S., and elsewhere (Gold et al. 2008;FAO 2015). The species also has been used recently as a model to investigate anthropogenic impacts on marine systems, including studies assessing impacts of agricultural chemicals Fuiman 2005, 2006) and ocean acidification (Diaz-Gil et al. 2015). Following the Deepwater Horizon oil spill in the GoM in 2010, red drum also has been selected as a model to study physiological and genetic effects of acute oil exposure in a marine fish.
Here, we present a dense genetic linkage map for red drum, combining 481 previously mapped anonymous and gene-linked microsatellite loci  with 1794 haplotyped contigs derived from double-digest restriction-site associated DNA (ddRAD) sequencing. Using the high degree of shared synteny observed between red drum and six other fish species whose genomes are published, we demonstrate how synteny data can be used to infer likely positions of candidate genes, in this case identified from transcripts obtained during a preliminary hydrocarbon exposure study, which could not be mapped using traditional linkage approaches.

MATERIALS AND METHODS
Prior to analysis of genetic data, a reduced-representation, reference genome for red drum was constructed. Twenty individuals from across the range of the species, including parents from two mapping crosses, were used to produce a ddRAD library following Peterson et al. (2012) as modified by Portnoy et al. (2015). The library was sequenced on an Illumina MiSeq DNA sequencer producing 300 bp, paired-end reads. Raw sequencing reads were demultiplexed using the program process_radtags from the Stacks package , and a de novo reference genome was assembled using the dDocent pipeline (Puritz et al. 2014). Because sampled RAD contigs had a mean size of 300 bp, the entire sequence of each RAD contig was recovered during reference assembly. A preliminary annotation of the reference genome with the BLAST algorithm revealed the presence of multi-copy nuclear ribosomal RNA (rRNA) genes. To avoid downstream issues with read mapping caused by multi-copy loci, a custom script was used to remove rRNA contigs, and all contigs with a total length of ,150 bp. All further analysis of RAD sequences utilized this reference genome for read mapping and SNP calling.
Tissue samples from two outbred, full-sibling mapping crosses (Family A: n = 117; Family B: n = 116) used to generate the microsatellite-based linkage map , were extracted using Mag-Bind Tissue DNA kits (Omega Bio-Tek). RAD libraries were constructed following procedures outlined in Portnoy et al. (2015), and sequenced on two lanes of an Illumina HiSeq 2000 DNA sequencer. Raw sequencing reads were demultiplexed, using the program process_radtags, to produce a file containing raw reads for each individual. Read mapping and SNP calling were conducted for each family separately, using the dDocent pipeline and the reducedrepresentation reference genome. Raw SNP genotypes were filtered stringently using the VCFtools package (Danecek et al. 2011). First, individual genotypes called from ,10 reads were removed, followed by all loci with a mean Phred quality score of ,20. An iterative filtering process was then used to maximize the number of individuals and loci in the final dataset. Loci with .50% missing data were excluded, followed by individuals with a mean depth of less than five reads and .95% missing data. Next, loci and individuals with .25% missing data were excluded, and loci with a minor allele frequency ,0.05 were removed. The bash script dDocent_filters (https://github.com/jpuritz/ dDocent/tree/master/scripts) was used to remove loci based on numerous criteria, including mean read depth, ratio of quality to depth, strand representation, allelic balance in heterozygous individuals, and proper read pairing. Complex polymorphisms were then decomposed to individual SNP or indel loci, using the vcfallelicprimitives command in the vcflib package (https://github.com/vcflib/vcflib). Loci were then collapsed into haplotypes within each RAD contig, using the program rad_haplotyper (Willis et al. 2016; http://www.github.com/chollenbeck/ rad_haplotyper). Briefly, the program uses read alignments to record combinations of SNPs (haplotypes) present across paired-end reads. It also applies filters to flag loci that have an excess of haplotypes (potentially indicative of paralogous loci) or a deficit of haplotypes (potentially indicative of genotyping error), given the SNP genotypes. The program removed loci that were haplotyped successfully in ,75% of individuals, and kept indel loci when the indel was the only polymorphism on the contig; other indels were excluded from the analysis to avoid complications in haplotyping. The resulting file for each family contained one diploid genotype per individual for all remaining RAD contigs. After filtering, the dataset for Family A consisted of 72 progeny with genotypes at 786 RAD contigs, and the dataset for Family B consisted of 81 progeny with genotypes at 1340 RAD contigs.
RAD-based genotype data were combined with microsatellite genotype data obtained previously , and the resulting data file imported into JOINMAP (van Ooijen 2012). RAD loci were then added to previously defined linkage groups . To reduce the chance of incorrectly adding loci to existing linkage groups, an initially conservative LOD score of 9.0 was applied, followed by two more rounds of grouping at an LOD of 6.0 and 3.0. After assigning loci to linkage groups, groups of loci that exhibited an observed recombination rate of zero were identified, and only a single representative locus was retained for initial ordering. Tests for segregation distortion were carried out using chi-square goodness-offit tests, implemented in JOINMAP. Family-specific maps were generated by applying the multipoint maximum likelihood algorithm for outbred crosses (van Ooijen 2011), as implemented in JOINMAP. Marker order for shared loci was compared between families, and incongruities corrected by identifying and removing problematic loci, which most often displayed either segregation distortion or contained probable genotyping errors. Loci excluded initially because of lack of recombination with another locus were then added to family-specific maps by placing them at the same map position as the representative mapped locus. Family-specific maps were combined into a consensus map with the program MERGEMAP (Wu et al. 2011), using equal weights for both families. To investigate whether mapped loci were located in protein coding genes, loci in the consensus map were screened against the NCBI nonredundant nucleotide (nt) database, filtered to include only vertebrate sequences, using the NCBI BLAST+ suite (Camacho et al. 2009). Hits were considered a successful match if the e-value for the hit was ,1 · 10 210 .
A custom pipeline (https://github.com/chollenbeck/synteny_mapper), written in the Perl programming language, was used to compare the linkage map for red drum to the genomes of six fish species for which chromosome-level scaffolding was available. Genome assemblies were downloaded for stickleback (Gasterosteus aculatus; gasAcu1), Nile tilapia (Oreochromis niloticus; onil1.1), green spotted puffer (Tetraodon nigroviridis; tnig_v8), fugu (Takifugu rubripes; FUGU5), European seabass (Dicentrarchus labrax; dicLab v1.0c), and barramundi (Lates calcarifer; v3). The pipeline first screened all mapped red drum loci for which sequences were available (2252 of 2275) for matches (e-value ,1 · 10 210 ) to the genome of each comparison species, using the discontiguous megablast algorithm in the BLAST+ suite. Red drum loci with significant matches to multiple chromosomes within a comparison species were discarded to prevent ambiguities related to multi-copy loci. For each locus that matched a single chromosome, the start position of the sequence on the chromosome was recorded. Next, regions of shared synteny, defined as regions consisting of at least two loci that shared a common marker order (collinearity) between the red drum linkage map and the genome of a comparison species, and which are uninterrupted by other mapped loci, were identified. The algorithm used to identify regions of shared synteny takes into account the possibility of small errors in genome assembly and map order, and allows small differences in marker order. In this case, departures from collinearity between loci separated by ,2.5% of the total linkage group or chromosome length were tolerated.
By using regions of shared synteny, it is possible to infer positions of unmapped loci, provided that a sequence for the locus is available, and it matches unambiguously to the genome of a comparison species . To demonstrate the utility of the red drum linkage map as a tool for inferring the genomic position of loci of interest, this approach was applied to an unpublished dataset of candidate genes identified from red drum transcripts obtained in a preliminary hydrocarbon exposure study (dataset doi: 10.7266/ N73T9F7J). Using the final module of the synteny_mapper pipeline, cDNA sequences of 724 gene transcripts were screened against the genome of each comparison species, using BLAST+, as above. Transcript sequences with single hits to the genome of at least one comparison species were identified, and for each locus it was determined whether the locus was located in a region of shared synteny between the red drum linkage map and the genome of the comparison species. When transcript sequences were located in a region of shared synteny and the region contained more than two mapped loci, the script identified the smallest possible interval within the region into which the locus could be positioned and recorded the nearest left and right flanking loci. The distribution of shared syntenic blocks and inferred positions of synteny-mapped loci, relative to the red drum linkage map, were visualized using ggplot2 (Wickham 2009) and circos (Krzywinski et al. 2009).

RESULTS
Following assembly, the reduced-representation reference genome contained 38,887 RAD contigs, with a mean size of 264.9 bp. Additional filtering for contigs below minimum threshold size (,150 bp), and those containing rRNA sequences resulted in a final reference assembly of 33,865 RAD contigs, with a mean size of 284.6 bp. The total length of the reference sequence was 9,638,003 bp. Assuming a total red drum genome size of 810 Mb (Gold et al. 1988), the reference covered 1.19% of the genome. After filtering, the dataset for Family A (72 progeny) consisted of 786 RAD contigs (containing 1383 SNPs), and the dataset for Family B (81 progeny) consisted of 1340 RAD contigs (containing 2620 SNPs). The difference in number of usable RAD contigs between families was largely the result of differences in overall DNA quality of individual samples from the two families. A detailed summary of filtering results is presented in Table S1. After combining RAD contig data with microsatellite and EST-SSR genotypes assayed previously from the same individuals, the total mapping dataset consisted of 1218 and 1779 loci for Families A and B, respectively.
The consensus linkage map (Figure 1) contained 2275 loci, including 1794 RAD contigs (consisting of 3462 SNP loci), 437 anonymous microsatellite loci, and 44 EST-SSRs; the combined length of mapped fragments totaled 692.3 kb. The mean number of loci per linkage group was 94.79 and the mean marker interval was 0.94 cM. The average length of a linkage group was 87.72 cM; the total map length was 2105.30 cM. Because of the tendency of MERGEMAP to inflate the total size of linkage groups when combining maps (Khan et al. 2012), the average of individual-specific maps may represent a more accurate estimate of total map size. Averaged across individual-specific maps, the mean length of linkage groups and total map length were 71.81 and 1704.40 cM, respectively. Of 2275 total loci, 278 (12.2%) were located directly in coding regions and could be assigned a putative function based on a BLAST search. Summary statistics for the consensus-, family-, and individualspecific maps are presented in Table 1; detailed information regarding map positions and annotations for the consensus linkage map are available in Table S2.
The total number of loci with significant BLAST hits to single chromosomes ranged from 342 (15.1% of all loci; green spotted puffer) to 1249 (55.3% of all loci; European seabass). The total number of blocks of shared synteny also varied among comparison species, ranging from 87 (green spotted puffer) to 284 (European seabass). The mean number of loci per block of shared synteny was less variable, ranging from 3.047 (stickleback) to 3.196 (fugu). Similarly, while the total size of blocks of shared synteny was highly variable among comparison species, ranging from n  92.9 Mb (green spotted puffer) to 259 Mb (Nile tilapia), the proportion of the genome of the comparison species covered by blocks was less variable, ranging from 0.330 (green spotted puffer) to 0.438 (European seabass). Summary statistics for blocks of shared synteny for each of the six comparison species are presented in Table 2; a summary of shared syntenic regions in the form of a circular ideogram displaying blocks for each species relative to red drum linkage groups is presented in Figure 2.
Of 724 candidate gene transcripts identified in oil exposure experiments, 227 (31.4%) were assigned a putative position, using the synteny-mapping strategy (Table S3). Of 277 synteny-mapped loci, 62 (27.3%) were annotated by a BLAST search. Inferred positions of all synteny-mapped genes are presented in Figure 2. The distribution of shared syntenic blocks and synteny-mapped genes on red drum linkage group 22, where eight candidate gene transcripts were localized to a region of 10 cM, are shown in Figure 3.

DISCUSSION
In total, 1794 haplotyped RAD contigs, comprising 3462 SNP loci, were added to the red drum linkage map. Addition of these loci reduced the mean marker interval to ,1 cM. The total length of the consensus map (2105.30 cM) is larger than reported previously for the microsatellite-based map (1815.3 cM, Hollenbeck et al. 2015); this difference is likely a result of combining maps from different families, using MERGEMAP, which is reported to inflate total map length (Khan et al. 2012). Relative to the average total length of individual-based maps reported here, the total length of the consensus map was inflated by 24%, consistent with the results of McKinney et al. (2015), who found that MERGEMAP increased the size of the Chinook salmon consensus map by 30% per additional family. However, there exists a tradeoff between accuracy of map lengths and inclusion of loci on the map . Because the purpose of our study was to determine the relative positions of loci, rather than the exact frequency of recombination between them, we chose to add more loci to the consensus map by combining information between mapping families.
The RAD-seq linkage mapping process employed here has a number of benefits. First, using long reads from the Illumina MiSeq platform enabled recovery of the entire sequence of each discrete ddRAD contig, thus providing more sequence coverage for comparative genomics and other downstream applications. Second, using haplotypes instead of individual SNPs eliminated redundancy and saved computational time by condensing tightly linked SNPs into single, multi-SNP loci, potentially possessing multiple alleles, thus discriminating more discrete alleles than would otherwise be identified, and, thereby, increasing the number of alleles that segregate in an informative manner (Ball et al. 2010). Last, haplotyping allows possible multi-copy loci to be filtered from the dataset because multi-copy loci often display more than two haplotypes within an individual and can be identified on this basis.
Of all comparison species, European seabass, followed by barramundi, had the highest degree of similarity with red drum in terms of total number of homologous loci (established via BLAST search), number and proportion of the genome contained in shared-synteny blocks. The lowest degree of similarity in all three measures was found between red drum and green spotted puffer. Correspondingly, shared-synteny block size was smallest in seabass and barramundi, and largest in fugu and green spotted pufferfish. Finally, in contrast to the considerable variation in the total number of BLAST hits and number of shared syntenic blocks among comparison species, the mean number of loci per block, and the proportion of the comparison species genome in blocks of shared synteny were much more similar among comparisons. There are a number of factors that likely produce this pattern, including the phylogenetic relationship between species, structural changes during genome evolution, and the quality of genome assemblies used in the analysis. European seabass, for example, is more closely related phylogenetically to red drum than green spotted pufferfish (Nelson et al. 2016), which likely explains the larger number of BLAST hits in European seabass (1249) as compared to the green spotted pufferfish (342). Having more homologous loci affords the opportunity to detect smaller blocks of shared synteny, as well as the fine-scale rearrangements and translocations that tend to break up larger regions of shared synteny into smaller blocks .
Results of the synteny analyses were consistent with a high degree of stability in karyotype evolution among teleost fishes (Amores et al. 2014). An important implication of this high level of shared synteny Figure 3 Shared syntenic blocks and synteny-mapped genes on red drum linkage group 22. Colored rectangles represent blocks of shared synteny between the red drum genetic map (horizontal axis) and genomes of six comparison species. Dark-red line segments represent regions where differentially expressed genes in oil exposure experiments were localized via a synteny-based mapping approach. Asterisks denote locations of genes that were synteny-mapped to zero-recombination intervals on the linkage map.
is that the information from well characterized genomes can be used to localize candidate genes of interest in nonmodel species. In our study for example, putative locations of previously unmapped candidate gene transcripts were achieved with a success rate of 31.4%, suggesting that the general approach will be useful in localizing candidate genes from environmental impact studies, or other gene expression studies in species for which a complete reference genome is unavailable.
The linkage map described herein will be a valuable resource for studies of red drum and other nonmodel species. Red drum, for example, are cultured both for restoration enhancement and commercial production purposes (Gold et al. 2008;FAO 2015). A linkage map can be used to great advantage for commercial aquaculture by enabling the mapping of quantitative traits and facilitating marker-assisted selection for genetic improvement (Liu and Cordes 2004), including the identification of chromosomal regions impacting disease resistance (Houston et al. 2012) and sex determination (Palaiokostas et al. 2013a,b). In addition, a combination of linkage and linkage disequilibrium (LD) data also has been shown to be effective in detecting changes in contemporary effective population size of wild red drum over time . Highdensity linkage maps also can be useful tools for chromosome-level scaffolding in genome assembly (Fierst 2015), and linkage map/ genome integrations are increasingly common (Gonen et al. 2014;Tine et al. 2014). Finally, linkage maps, when combined with population genetics data, can provide a genomic context for loci used in population-level studies to identify genomic islands of adaptation (Bradbury et al. 2013) and associations among loci implicated in adaptation .