Hormonal sex reversal and neomale verification

RAD sequencing

Two crosses with 62 (Family C) and 28 offspring (Family B) and their parents (including a common female) were analysed (Additional file 1). The DNA samples were barcoded, pooled and sequenced in two lanes of an Illuminia HiSeq 2000 sequencer (see Methods). In total, 648,402,546 raw reads (101 bases long) were produced (or 324,201,273 paired-ended reads: NCBI BioProject accession number SRP016043). After removing low quality sequences (quality score under 30), ambiguous barcodes and orphaned paired-end reads, 81.24% of the raw reads were retained (526,783,920 reads). The Stacks package [25] was then used to make a de-novo assembly of the sampled loci from each individual: 82,745 and 83,668 RAD-tags were retrieved for Families B and C respectively, covering 90,105 RAD-tags in total including 76,308 of these shared between the two families (Figure 1). The number of reads and RAD-tags for each sample are reported in Additional file 1.

Genetic map

In order to maximise the number of informative markers and minimise the amount of missing or erroneous data, we used only paired-end RAD-tags retrieved in at least 75% of the samples in each family, and carrying one or two SNPs. 7572 and 5954 RAD-tags were retained for the families B, and C respectively (Figure 1). Since Family B had only 28 offspring, the genetic map was constructed with the Family C data only (62 offspring). The map consists of 5703 SNPs and 7 microsatellites (used initially for parentage assignment) in 24 linkage groups (LG) and spanning 1514 cM (Figure 2; Additional file 2). 4049 of the above SNPs were common in the two families and were used to incorporate the data from Family B into a joint linkage map. Sex-specific genetic maps were constructed, with the female-specific map spanning 1496 cM and the male-specific map spanning 1378 cM (Table 2).

QTL-association mapping

The results from the single-QTL model for binary traits provided evidence for the existence of a major QTL in LG 13 for both families in the male informative dataset (Figure 3A). The highest logarithm of odds (LOD) scores for Families B and C were 6.83 and 12.17 respectively (Figure 3B). The genome-wide thresholds (α = 0.01) were calculated from permutation tests (10,000 permutations) to be 3.25 and 3.90 respectively. The highest LOD scores were observed in the region between 30 cM and 48 cM in LG 13 for both families. Even though models that take into account the existence of a major QTL (as in this study) or ones that test for existence of multiple QTLs simultaneously reduce the residual variation (providing this way higher power in the analysis for detecting additional QTLs at least of modest effect), no additional QTLs were detected. The calculated 95% Bayesian Density Intervals for the QTL location spanned a region of 22 cM (28-50 cM in LG 13), while the 1.5-LOD support interval spanned a region of 28 cM (23-51 cM in LG 13). The variance explained by the QTL was around 82%. There were no significant QTL in the female informative dataset. The Polymorphism Information Content (PIC) of the markers in LG 13 ranged from 0.64 to 0.99. The joint QTL analysis of the two families showed a maximum at 46 cM with a Likelihood Ratio (LR) of 96 (Figure 3B). The chromosome-wide threshold calculated from permutation tests (10,000 permutations) was LR = 19. The association analysis identified 38 SNPs being strongly associated (P < 0.001) with sex in the two families. Models scanning simultaneously for the existence of two QTL or multiple QTL did not reveal any extra QTL.

Verification of SNP sex association and sex prediction

Sex association of 10 SNP markers (Additional file 3) selected from LG 13 was investigated using allele specific endpoint-genotyping assays (Figure 4; Additional file 3). The P-value thresholds (α = 0.05) after taking into account multiple testing (for the 10 SNP markers that were tested) according to the permutations and the Bonferroni correction tests were found to be 0.0054 and 0.0050 respectively. Hhi 58665 (P = 3.05 × 10^-19), Hhi 10170 (P = 3.05 × 10^-19), Hhi 41238 (P = 2.45 × 10^-16) and Hhi 47769 (P = 8.96 × 10^-17) showed the highest association with sex, over 97% in the three test families (n = 10 ♀:10 ♂ in Families 1 & 2; 9 ♀:9 ♂ in Family 3; total 58). In each case the parents were a heterozygous ♂ and a homozygous ♀ for each marker. Furthermore when tested in 36 wild sourced broodstock from locations across the species native range (Additional file 4), the same four markers individually showed between 78% and 89% association with sex (P < 0.001), with SNP Hhi 58665 showing the highest association (P = 6.39 × 10^-7). These four markers were located within a 3.2 cM region of LG 13.

The combined prediction power of these four markers was tested on the 36 broodstock and 58 progeny using the JRip classifier, as the derived rules have straightforward interpretation. The combined prediction is based on two rules using three markers (Figure 5) and produced 97% corrected classification with the 3% of errors being phenotypic males predicted to be females. The male prediction precision is 1 (recall of 0.935) and the female prediction precision is 0.94 (recall of 1).

Within the 18 tested male broodstock, two had “female” genotypes for all four of these markers (one of the 58 progeny also had male phenotype but female genotype). One of these two broodstock had previously been crossed with four females and had produced only female offspring (between 3 and 14 individuals per family, total 27). The phenotypic sex of these 27 offspring was verified by post-mortem examination three years post-fertilisation.

Synteny searches

We selected the 59 markers within the 95% confidence interval around the LOD score peak and mapped them onto the genomes of related species to identify syntenic regions. We performed this search against Danio rerio (zebrafish), Gadus morhua (Atlantic cod), Gasterosteus aculeatus (three-spined stickleback), Latimeria chalumnae (West Indian ocean coelacanth), Oreochromis niloticus (Nile tilapia), Oryzias latipes (medaka), Takifugu rubripes (tiger pufferfish) and Tetraodon nigroviridis (spotted green pufferfish) genomes. 33 markers had unique hits across at least five out of eight species (Additional file 5). G. aculeatus and O. niloticus show the highest level of synteny with H. hippoglossus and each other (Additional file 5). The order of the markers selected for SNP genotyping in the regions point toward one 3.2 Mb region embedding more than 60 annotated genes (See direct links to the Ensembl 68 in Additional file 5). No genes associated with sexual differentiation or determination were identified in this region.

Ethics statement

All working procedures complied with the United Kingdom Animals (Scientific Procedures) Act 1986 and were approved by the ethics committee of the University of Stirling.

Hormonal sex reversal

Weaned mixed-sex halibut larvae (mean total length of 40.1 ± 0.2 mm, mean wet weight of 0.5 ± 0.01 g) produced in the 2007 spawning season, were obtained from a commercial halibut hatchery and transferred to the Machrihanish Marine Environmental Research Laboratory (55.424°N, 5.749°W) for hormonal treatment. Three in-feed treatments were tested in duplicate: a) 6 weeks steroid free diet (control), b) 6 weeks MDHT in-feed (5 ppm) [18] and c) 3 weeks MDHT in-feed (10 ppm) followed by 3 weeks steroid-free diet. Food was provided in excess by automated feeders into the tanks every 12 minutes throughout a 24-hour period. Feed, based on a commercial diet (Low Energy Marine Larval diet, EWOS, West Lothian, UK), was mixed with an ethanol solution containing the appropriate dose of MDHT (Sigma-Aldrich Co Ltd, Poole, UK) and then dried in an extraction fume hood. Following treatment and once the fish had reached a mean weight of 28.4 ± 0.4 g, replicate treatment groups were identified by a coded subcutaneous dye mark and then reared communally. At approximately 1 year post fertilisation a total of 80 individuals per treatment (40 per replicate) were sacrificed and fixed in 4% neutrally buffered formalin for histological determination of phenotypic sex. Sex ratios were compared to the expected 1:1 and were evaluated statistically using a Chi-square test, (P < 0.05, χ² = 3.84, df = 1).

Neomale verification by progeny testing

At a mean size of 180.8 ± 3.1 g, 60 control fish (30 per replicate) and 150 fish from the 5 ppm treatment (75 per replicate) were tagged with a passive integrated transponder tag (Fish Eagle Co., Lechlade, UK). Fish were then reared communally until first maturity in spring 2010. In March 2010, crosses were performed between 7 males from the hormone-treated population and normal female broodstock. Fertilisation was confirmed in each cross by microscopic examination of blastomere development. Eggs from each cross were maintained in isolation using standard commercial rearing methodologies. Sufficient progeny from only four of these males survived through yolk sac absorption, live feeding and weaning. These four families were reared in isolation at a commercial halibut hatchery until phenotypic sex ratio could be assessed in February-March 2011, once fish reached a suitable size (over 50 g) for histological sexing of the gonads. A total of 30 (family A & B) or 70 (family C & D) individuals/family were sacrificed for histological examination and blood was sampled for genotyping (total of 200 offspring). Sex ratios were compared to the expected 1:1 using a Chi-square test (P < 0.05, χ² = 3.84, df = 1).

RAD library preparation and sequencing

DNA was extracted from blood samples of the fish using the REALPure genomic DNA extraction kit (Durviz S.L.) and treated with RNase to remove residual RNA from the sample. Each sample was quantified by spectrophotometry (Nanodrop) and quality assessed by agarose gel electrophoresis, and was finally diluted to a concentration of 50 ng/μL in 5 mmol/L Tris, pH 8.5. The RAD library preparation protocol followed essentially the methodology originally described in Baird et al. [23] and comprehensively detailed in Etter et al. [40], with the minor modifications described in Houston et al. [38]. The RAD-specific P1 and P2 paired-end adapters and library amplification PCR primer sequences used in this study are detailed in Baxter et al. [41].

Each sample (1.5 μg parental DNA / 0.5 μg offspring DNA) was digested at 37°C for 30 minutes with Sbf I (recognising the CCTGCA|GG motif) high fidelity restriction enzyme (New England Biolabs; NEB) using 6U Sbf I per μg genomic DNA in 1× Reaction Buffer 4 (NEB) at a final concentration of c. 1 μg DNA per 50 μL reaction volume. The reactions (75 / 25 μL final volumes for parental / offspring samples respectively) were then heat inactivated at 65°C for 20 minutes. Individual specific P1 adapters, each with a unique 5 bp barcode (Table 1), were ligated to the Sbf I digested DNA at 22°C for 45 minutes by adding 3.75 / 1.25 μL 100 nmol/L P1 adapter, 0.9 / 0.3 μL 100 mmol/L rATP (Promega), 1.5 / 0.5 μL 10× Reaction Buffer 2 (NEB), 0.75 / 0.25 μL T4 ligase (NEB, 2 M U/mL) and reaction volumes made up to 90 / 30 μL with nuclease-free water for each parental / offspring sample. Following heat inactivation at 65°C for 20 minutes, the ligation reactions were slowly cooled to room temperature (over 1 hour) then combined in appropriate multiplex pools (Additional file 1). Shearing (Covaris S2 sonication) and initial size selection (250-500 bp) by agarose gel separation [38] was followed by gel purification, end repair, dA overhang addition, P2 paired-end adapter ligation, library amplification, exactly as in the original RAD protocol [23, 40]. A total of 150 μL of each amplified library (14 PCR cycles) was size selected (c. 300-550 bp) by gel electrophoresis [38]. Following a final gel elution step into 20 μL EB buffer (MinElute Gel Purification Kit, Qiagen), the libraries were sequenced at The GenePool Genomics Facility at the University of Edinburgh, UK, for quality control and high-throughput sequencing. Libraries were accurately quantified by qPCR (Kapa Library) and run in two lanes of an Illumina HiSeq 2000 using 100 base paired-end reads (v3 chemistry). Raw reads were process using RTA 1.12.4.2 and Casava 1.6 (Illumina). The reads were deposited at the NCBI BioProject under the accession SRP016043.

Genotyping RAD alleles

Reads of low quality (score under 30, while the average quality score was 37), missing the restriction site or with ambiguous barcodes were discarded. Retained reads were sorted into loci and genotyped using Stacks software 0.9995 [25]. The likelihood-based SNP calling algorithm [42] implemented in Stacks evaluates each nucleotide position in every RAD-tag of all individuals, thereby differentiating true SNPs from sequencing errors. The parameters were a minimum stack depth of at least 30, a maximum of 2 mismatches allowed in a locus in an individual and up to 1 mismatch between alleles. The pair-ends were assembled using Stacks and Velvet version 1.2.08 [43] and used to separate RAD-tag sequence with or without potential SNP but belonging to separate loci (duplication products). Polymorphic RAD-tags may contain more than one SNP, but the vast majority (over 99%) showed only two allelic versions; the very small proportion of RAD-tags with more than two alleles were excluded.

Genetic map construction

The genetic map was constructed using R/Onemap [44] and TMAP [45]. The allocation of markers in linkage groups was conducted using R/Onemap. This package uses Hidden Markov Models (HMM) algorithms for outbred species while in parallel implements the methodology described in Wu et al. [46] for calculating the most probable linkage phase. Linkage groups were formed using minimum LOD values of 10. TMAP was used to order the markers in every linkage group. By using an HMM maximum likelihood model and taking into account potential genotypic errors it reduces the tendency to erroneously derive oversized linkage groups, a phenomenon which is often observed in dense maps [45]. Map distances were calculated in centiMorgans (cM) using the Kosambi mapping function. The genetic map was drawn and aligned using Genetic-Mapper v0.3 [47].

QTL association mapping

The QTL analysis was performed using three different suites of programmes: R/qtl [48], GridQTL [49] and QtlMap [50]. In the case of R/qtl the genotypes the two families were analysed separately. The analysis was performed considering the cross as a ‘pseudo’ backcross, effectively analysing male and female informative markers separately. The model used for the analysis was based on Interval Mapping. The phenotype was considered a binary trait (0 for females and 1 for males). The algorithm used considers the phenotype to follow a mixture of Bernoulli distributions and uses a form of the EM algorithm for obtaining maximum likelihood estimates [48]. Two-way and multiple QTL models were also run with this package. Approximate Bayesian and 1.5-LOD 95% density and confidence intervals were calculated respectively. An approximate estimate of the phenotypic variance explained by the QTL was obtained from the following equation: 1-10^-2LOD/n. While the estimated variance may be reasonable for additive QTL, problems can be caused in the case of linked QTL [48]. The GridQTL software was used to estimate the polymorphism information content across the genetic map. QTLMap was used for performing a joint QTL Analysis of the two families. The phenotype was considered as discrete and the model used was a Mixture Linkage Analysis model, accounting for heteroskedasticity. An Association Analysis was performed for the two families using R/GenABEL [51] in order to identify SNPs associated with sex. The SNP data were tested for association using the fast score test for association [52]. In all the above analysis genome-wide significance thresholds were calculated by permutation tests (10,000 permutations) in order to correct for multiple testing.

Verification of SNP sex association

Marker sex association was tested using 10 competitive fluorescent, allele specific endpoint-genotyping assays (KASP v4.0, LGC genomics) based on SNPs that were commonly found in the two mapping families to span the region of highest association with sex (Hhi6696, Hhi7153, Hhi9493, Hhi10170, Hhi11772, Hhi18571, Hhi41238, Hhi47769, Hhi51454, Hhi58665, NCBI dbSNP accession 749737483, 749737484, 749737485, 749737486, 749737487, 749737488, 749737489, 749737490, 749737491 and 749737492 respectively; Additional file 3). SNP-specific primer sets were designed by LGC genomics (Additional file 3). Each genotyping assay was run in an 8 μl volume containing approximately 40 ng of target gDNA incorporated with a proprietary reaction mix in accordance with the manufacturer’s guidelines. All assays were run using the same touchdown thermal cycling programme as follows: 94°C for 15 minutes followed by 10 cycles of 94°C for 20 seconds melt, 65-57°C for 1 minute anneal and extension (decreasing of 0.8°C per cycle) followed by 26 cycles of 94°C for 20 seconds melt, 57°C for 1 minute anneal and extension. There was one exception, SNP Hhi 58665, for which the extension time was extended to 2 minutes. All assays were run in a Biometra TGradient thermal cycler (Biometra GmbH, Goettingen, Germany). Thereafter assays results were read at 25°C using an endpoint genotyping programme in a Techne Quantica qPCR thermal cycler (Bibby Scientific Ltd, Stone, UK) in which unknown genotypes were assigned based on fluorescent output in comparison to non-template control wells containing DNA/RNA free H₂O. All 10 SNP assays were tested in 58 offspring from three halibut families produced in the commercial halibut hatchery, which were independent from the initial mapping families, and in 36 independent broodstock halibut (18 ♀:18 ♂) originating from the Shetland Isles, Iceland and possibly the Faroe Islands. An association analysis was performed using R/SNPassoc [53]. In the case of family data, association was tested both in separate families and across all families together. A Bernoulli generalised linear model was applied in order to test the magnitude of association between the SNP genotypes and phenotypic sex using this package (function association). Both the Bonferroni and permutation tests (10,000 permutations) were used in order to correct for multiple testing.

Sex prediction

The KASP allele type of all markers for each individual tested along with their sex were entered into the WEKA package [54], which contains a variety of machine-learning algorithms, including JRip, an optimised rule learning algorithm. This classifier implements a propositional rule learner, Repeated Incremental Pruning to Produce Error Reduction (RIPPER), which was proposed by Cohen [55] as an optimised version of IREP. JRip builds additive rules based on the allele type of the markers. JRip then classifies each individual into a particular predicted sex based on the allele type of the markers for each individual. Permutatively, one individual was removed from the training set, and subsequently the algorithm then assigns its sex. The set of rules was stable between permutations (Figure 5).

Synteny searches

D. rerio, G. morhua, G. aculeatus, L. chalumnae, O. niloticus, O. latipes, T. rubripes and T. nigroviridis genomes were downloaded from Ensembl 68 [56]. We used BLASTN [57] to perform a search for the RAD-tag (and their paired-ends) against the 8 fish genomes. The parameters used were minimum alignment size 80 nt, minimum percentage of sequence identity 0.25 and maximum e-value 0.001 and low complexity mask on. All other parameters were set as default to account for the divergence and shortness for the sequences used. Sequences that aligned to more than one place in each genome were excluded from further analysis.