A total of 17,056 EST sequences were aligned and used to detect SNPs. The ESTs were 5' sequences representing eight different cDNA libraries with five to several hundred animals per library. The average sequence length was 508 ± 137 bp (SD). The alignment of ESTs produced 1186 contigs and 207 single sequences (Table 1). Seven hundred and twenty-four putative SNPs were found, distributed on 415 contigs. The contigs having candidate SNPs consisted of two to 11 EST sequences.

Assays were successfully developed for 648 of the 724 putative SNPs. Five hundred and ninety-four of these 648 SNPs were genotyped on a collection of Atlantic cod sampled at four different locations; one site populated primarily by NEAC, one northern (NCC-N) and two southern (NCC-S1 and NCC-S2) sites populated primarily by NCC (Figure 1, Table 2). NCC-S1 and NCC-S2 were found not to be significantly different from one another (pairwise F_ST = 0.013, p-value = 0.074), and are hereafter treated as one population, termed NCC-S. The animals that were genotyped had been preselected from larger samples based on their Pan I genotype. Thus, all NEAC were Pan I^BB and all NCC were Pan I^AA.

Twenty-nine percent of the genotyped SNPs failed to amplify or did not cluster well according to genotype and were considered "failed assays", 2% were heterozygous in all animals and were thus assumed to represent duplicated genes, 15% were homozygous in all animals, while 54% were polymorphic and reliably scored (Table 3). The 318 polymorphic SNPs represented 235 contigs, with one to five SNPs per contig (Table 3). The mean minor allele frequency among the polymorphic SNPs was 0.18 ± 0.15 (SD) (Figure 2), while the mean overall heterozygosity was 0.25 ± 0.17 (SD). When homozygous SNPs were included, the heterozygosity was 0.19 ± 0.18 (SD). The true heterozygosity of real SNPs in the data set is likely to lie somewhere in between these two estimates, since we could not distinguish real SNPs that were homozygous in our material from loci that had falsely been identified as putative SNPs.

Annotation using BLASTN and BLASTX gave significant hits for all but 73 of the 594 genotyped SNPs. Synonymous/non-synonymous status of SNPs could be determined for 77 of 594 SNPs; the remaining 517 were either located outside coding regions, or the corresponding genes had not been annotated. The overall ratio of synonymous to non-synonymous SNPs was 0.44, which agrees well with results from other species. For example, Cargill et al. 27 and Lee et al. 28 reported the synonymous-to-total ratio to be 0.47 and 0.45 in humans and cattle, respectively. The annotations of those contigs that contained true SNPs are given in Additional file 1.

Since animals had been assigned to populations based on genotypes at Pan I, we wanted to test for animals that might have been misclassified as NEAC, NCC-N, or NCC-S. The assignment test reassigned three animals in the NEAC sample to NCC-N, and one animal in NCC-N to NEAC. These four animals were culled from further analysis. The three misclassified animals from the NEAC sample had caused significant deviations from Hardy-Weinberg equilibrium (HWE) at individual markers with large F_ST-values, indicating that a few high-F_ST SNPs could be sufficient to substantially improve the accuracy of NEAC/NCC assignments compared to assignment with Pan I alone.

After culling of misclassified animals, the fraction of polymorphic SNPs that were not in Hardy-Weinberg Equilibrium (HWE) (P < 0.05) was 0.10 for NEAC, 0.04 for NCC-N, and 0.09 for NCC-S, i.e. slightly higher than expected by chance. Four SNPs, Gm376_0668, Gm0927_0134, Gm0786_0527, and Gm392_1244, were out of HWE in all three populations, while 12 other SNPs were out of HWE in two populations [see Additional file 1]. Deviations from HWE could be due to genotyping errors or bias in the identities of uncalled genotypes. However, genotypes of SNPs out of HWE in one or more populations had been carefully double-checked, and the fraction of markers out of HWE were not changed significantly by removing markers with call rates < 0.90, indicating that genotyping error is not likely to be the only cause of deviations from HWE in the data set.

We wanted to investigate the potential of the SNP markers for discriminating different populations of Atlantic cod, in particular NEAC versus NCC, using the SNP genotypes described above. The mean pair-wise F_ST-values were 0.057 and 0.024 for NEAC – NCC-N and NCC-N – NCC-S respectively, both significantly different from zero (P < 0.05). At individual markers, F_ST-values ranged from 0 to 0.83 (Figure 3), and 48 SNPs had F_ST-values significantly different from zero [see Additional file 1]. The Beaumont and Nichols test 29 revealed 29 outliers from the neutral-model F_ST distribution (Figure 3, Figure 4). The majority of these 29 SNPs had an allele frequency cline reminiscent of that observed at the Pan I locus [see Additional file 2]. The outlying SNPs represented 26 genes, of which five corresponded to ribosomal proteins, three were involved in muscle contraction, two were involved in immune response, seven had other known functions, and nine had unknown or ambiguous functions [see Additional file 2].

We did not aim to thoroughly investigate the complicated issue of Atlantic cod population structure in this study, given the limited number of populations investigated and the pre-selection of individuals according to their Pan I genotype. However, since very large F_ST-values were observed, it seems reasonable to conclude that outlier loci are, or have been, subject to diversifying selection, and that the gene flow between populations is minor; allele frequency divergence at the scale observed here would else not be likely to occur. Selection could be acting i) on the outlier SNPs themselves or on other SNPs within the same gene, or ii) on other loci in linkage disequilibrium with the outlier SNPs. Given the large number of genes in any higher organism, the even larger number of SNPs, and the limited number of markers tested in this study, selection is quite likely to be acting on genes in LD with the observed genes rather than on the observed genes themselves. Determination of LD levels in Atlantic cod will thus be essential to differentiate these two possibilities. If LD can be found across large genomics regions, the outlier genes/SNPs detected here are not likely to be under direct selection. However, if LD can be found across only minor genomic regions, selection is likely to be acting directly on the genes/SNPs themselves, in which case the data indicate that as much as 11% (26 out of 235) of genes in Atlantic cod are under positive selection.

EST sequences in Atlantic cod were obtained from eight different cDNA-libraries, made from tissues collected from the head kidney, intestine, liver, pyloric saeca, spleen (two libraries; stimulated and non-stimulated), and embryos (two libraries). For the embryo libraries, 500 different embryos were used, while for the other libraries, 15 animals were used in total. The animals used for EST sequencing were all Norwegian coastal cod. Samples from individual fish were stored in RNALater (Ambion) prior to use. Total RNA was extracted from each individual separately by the RNA Easy Mini Kit (Qiagen, Venlo, the Netherlands) and an equal amount of RNA from five different individuals was pooled. Enrichment for polyA RNA was conducted using the Oligotex mRNA Midi Kit from Qiagen. Nucleic acid quality was measure using Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, USA) and quantity was measure using Nanodrop ND-1000 (NanoDrop Technologies, Wilmington, USA). Directional cDNA libraries were constructed using 5 μg of polyA enriched RNA using the pBluescript II XR cDNA library construction kit from Stratagene (Cedar Creek, USA) according to the manufactures recommendations. The quality (size) (Bioanalyser) and quantity (Nanodrop) of the produced cDNA was assessed prior to ligation. XL10-gold cells were transformed and blue/white screened. Positive (white) colonies were randomly picked and grown overnight in 96 well plates. Plasmids were purified using the Montage Plasmid Miniprep kit (Millipore, Billerica, USA). Sequencing was conducted using BigDye chemistry and they were run on an ABI3700 sequencer from Applied Biosystems (Foster City, USA). Sequencing was done from the 5' end of transcripts.

Base-calling from chromatograms was performed using the Phred program 30. Vector (pCMV-PCR) was masked using Cross-Match (P. Green, unpublished; 31). PolyA and polyT sequences were also masked using a custom made script, to avoid false clustering on these motifs. Clustering and contig assembly was performed with the Phrap program, (P. Green, unpublished). Phrap was run with the options "-trim_start 50 ~minmatch 50. The PolyBayes program 32 was used to detect putative SNPs in the sequence alignments, and give a probability of being a true SNP to each base substitution. The prior polymorphism rate for PolyBayes was 0.001 In addition, all contigs were manually inspected for SNPs using the program Consed 33.

In order to annotate the SNP-containing contigs, BLASTX was run against the Protein data base, Swiss-Prot and non-redundant GenBank databases. A significant database hit was defined as having an expectation value (E-value) below 1.0 × 10^-10. All sequences with a significant BLASTX hit in Swiss-Prot were annotated by annotation transfer, inferring similarity of function from sequence similarity by applying the Gene Ontology (GO) assignments for the UNIPROT database, produced by the GOA project of the European Bioinformatics Institute 34. The gene_association goa_uniprot database of 07.07.2004 was used (see 35 for details) together with GO terms form the GO release of 06.08.2004 36. The sequences were annotated based on the single best hit in the Swiss-Prot database. BLASTN was used to annotate contigs against the nucleotide data base. Putative SNP functionality (e.g. resulting in a change in amino acid sequence or no change in amino acid sequence, non-synonymous or synonymous changes respectively) was predicted using the cSnper program 37.

The fish originated from randomly selected wild-caught Norwegian coastal cod (NCC) from southern and northern locations as well as from North-East Arctic cod (NEAC) (Table 2). All samples were collected autumn 2003. Genomic DNA was isolated from ethanol preserved fin clips in a GenoM-48 Robotic Workstation (Genovision, Oslo, Norway) using Magnetic Bead kits for purification and extraction of nucleic acids (Genovision) following the manufacturer's instructions. The fish were genotyped to characterise the fish as coastal cod or North-East Arctic cod based on their Pantophysin (Pan I genotypes 38. The fish used in this study were a subset of the samples taken at the locations, selected based on their Pan I genotypes, where Pan I^AA were regarded as NCC and Pan I^BB were regarded as NEAC (Table 2). Pan I genotyping was done according to Stenvik et al. 38.

Genotyping of SNPs was performed using the MassARRAY system from Sequenom (San Diego, USA). PCR-primers and extension-primers were designed using the software SpectroDESIGNER v3.0 (Sequenom). Multiplexes and primer sequences can be found in Additional file 1. Multiplexing levels were between 20 and 29. All SNP genotyping was performed according to the iPLEX protocol from Sequenom (available at 39). For allele separation, the Sequenom MassARRAY™ Analyzer (Autoflex mass spectrometer) was used. Genotypes were assigned in real time 40 by the MassARRAY SpectroTYPER RT v3.4 software (Sequenom) based on the mass peaks present. All results were manually inspected, using the MassARRAY TyperAnalyzer v3.3 software (Sequenom). SNPs were classified as "failed assays" (meaning that the majority of genotypes could not be scored and/or the samples did not cluster well according to genotype), "SNPs w/all animals heterozygous", "SNPs w/all animals homozygous", or "polymorphic SNPs", based on this manual inspection. SNPs that were out of HWE in one or several populations were double-checked.

Mean pairwise F_ST values were calculated using the "pairwise F_ST" function of Arlequin v3.11 41. F_ST values at individual SNPs were calculated using the AMOVA function of the same program. Permutation testing with 1000 iterations was used to calculate p-values for mean and locus-by-locus F_ST values. Arlequin v3.11 was also used for exact tests of HWE equilibrium (100 000 Monte Carlo iterations, 1000 dememorisation steps), and for assignment of individual genotypes to populations. Outlier SNPs were tested for using the method of Beaumont and Nichols 29, as implemented in the software package FDIST2 42. In FDIST2, average heterozygosity was first calculated using the datacal function. Simulations were then run using this average heterozygosity, 20 000 iterations, and assuming 10 demes, 3 populations, 40 individuals per sample, and a stepwise mutation model (the number of demes was increased until the average of simulated F_ST values was approximately equal to the observed average F_ST, as suggested in 29). Simulated F_ST values were plotted against heterozygosity to yield a distribution for F_ST under a neutral model. SNPs with F_ST values above the 0.975 quantile were considered to be outlier loci.