Domesticated Atlantic salmon can escape from aquaculture facilities into the wild, and in the period 2001-2005, 260 000-715 000 farmed escapees were officially reported in Norway to the Directorate of Fisheries annually. However, the true number of escapees is probably higher due to underreporting 30. In an attempt to assist the Norwegian authorities in improving regulation over the aquaculture industry, the Institute of Marine Research in Norway developed a DNA based forensic method to identify escaped salmon to farm of origin 2227. In short, this method is based upon screening a panel of STR loci on samples of escaped salmon that are recaptured in the wild in addition to salmon collected from farms in the surrounding area that are considered as potential sources of the escapees. A combination of genetic assignment in addition to probability based exclusion is used to identify the most likely source(s) of origin for the escapees.

The data set chosen for analysis in the present study consisted of approximately 500 Atlantic salmon resulting from investigating an unreported escapement episode. Fish were sampled from nine cages located on six marine farms (hereon referred to as samples A-I), in addition to 50 farmed escapees (RF) that had been recaptured in the vicinity of these farms. The recaptured fish were distinguished as farmed salmon based upon morphological characteristics. The farms, locations and exact dates of collection remain anonymous for legal reasons.

DNA extraction and STR analysis was performed at the Institute of Marine Research (IMR) in Bergen. DNA extraction was carried out in 96 well format using a Qiagen DNAeasy kit according to the manufacturers' instructions. Each plate contained a minimum of two negative controls. DNA was extracted twice for 48 of the 50 escapees (separate dates). The following fifteen STR loci were amplified in three separate multiplex PCR reactions; SSsp3016 (Genbank no. AY372820), SSsp2210, SSspG7, SSsp2201, SSsp1605, SSsp2216 31, Ssa197, Ssa171, Ssa202 32, SsaD157, SsaD486, SsaD144 33, Ssa289, Ssa14 34, SsaF43 35, using a modification of a previously described protocol 36. These loci are routinely used for performing Atlantic salmon genetics studies at IMR but do not represent an optimised set of loci for performing genetic assignment of farmed escapees. Locus SsaD486 was monomorphic in the entire data set and was excluded from all statistical analyses. PCR products were analysed on an ABI 3730 Genetic Analyser using the 500 LIZ™ size-standard. Alleles were automatically binned and manually checked in the Genotyper software prior to data analysis. A total of 87 individuals (from individual samples displaying partial PCR amplification failure on first amplification/electrophoresis) in addition to 48 of the 50 escapees were re-analysed (pcr amplification then electrophoresis). These individuals served as a genotyping controls. Several authors 637 have recommended the routine use of genotyping controls in genetic data sets to estimate error rates.

SNP genotyping was performed using the MassARRAY platform from Sequenom (San Diego, USA). A total of 388 SNPs were included in the study. Map position and flanking sequence of the majority of these SNPs are from Moen et al. 2 and Lorenz et al. 38 (Additional file 1). Multiplexes and primer sequences for genotyping are available upon request. All SNP genotyping was performed according to the iPLEX protocol (available at http://www.sequenom.com) using the MassARRAY™ Analyzer (Autoflex mass spectrometer) from Sequenom. Genotypes were assigned in real time 39 by using the MassARRAY SpectroTYPER RT v3.4 software followed by manual inspection of genotypes using the MassARRAY TyperAnalyzer v3.3 software.

In order to compare the two classes of markers, the STR and SNP data sets were mixed in various combinations. These are described in the results section as some combinations were test specific. However, the following four data sets were used as the start point for the majority of the analyses within: 1) 14 STR loci, 2) 300 SNP loci, 3) 28 SNP loci (selecting the SNP displaying the highest global F_ST for each of the 28 linkage groups), and 4) 195 mapped SNPs with minimal distance of > 1 cM to nearest SNP (selecting the SNP displaying the highest global F_ST for 2 or more SNPs < 1 cM). For some tests, loci were ranked prior to computation. Ranking of loci was carried out by three methods including number of alleles displayed through samples A-I (STR loci only), global F_ST across samples A-I (STR and SNP loci), and with the locus assignment power program BELS across samples A-I 40. BELS was programmed to maximise mean individual assignment power on the data set without any re-sampling.

Allelic variation, heterozygosity and F statistics were computed in the program MSA 41. Arlequin V3.11 42 was used to calculate deviation from Hardy Weinberg Equilibrium. MEGA 43 was used to produce phylogenetic trees for the various data sets using the UPGMA method on matrices of pair-wise F_ST values. The trees were linearised assuming equal evolutionary rates in all lineages 44. Geneclass V. 1.02 45 was used to perform self-assignment simulations among samples A-I using the leave one out sub-option, and direct assignment of the escapees to these samples. All tests were performed using the Rannala and Mountain 46 method of estimating allele frequencies unless otherwise stated.

Bayesian clustering analysis implemented in STRUCTURE 2.2 4748 was used for estimating the number of populations/groups (k) represented by the data set. Following pilot analysis, main runs assuming k = 1-6, each with 3-5 iterations (depending upon data set), were conducted in order to estimate the most likely k and to assign individuals to these groups without using prior information about their sample of origin using correlated allele frequencies and an admixture model. Each run consisted of a burn-in of 100,000 MCMC steps, followed by 500,000 replications. STRUCTURE 2.2 was also used to perform a modified self-assignment procedure by removing 10 individual salmon at random from samples A, D, F and G, then assigning these individuals to the baseline which did not include those individuals. This was conducted by using the prior population information for the baseline samples and no prior population information for these 40 individuals. The results obtained from this analysis were compared to an identical procedure in Geneclass for these 40 individuals (in the latter case removing the 40 fish from the baseline file and entering them into a separate file of "unknown" individuals).

DNA was isolated from a total of 512 salmon that were analysed for 15 STR loci at IMR and 388 SNP loci at CIGENE. In the two laboratories, 13 and 14 of the individual DNA extracts failed to yield PCR products for any of the loci, leaving STR and SNP data sets consisting of 499 and 498 individual fish respectively. Individuals failing to yield PCR products for the STR loci were spread among samples, whereas all complete amplification failures in the SNP data set were observed within the sample of escapees.

Within the STR data set, a total of 87 individuals displaying PCR failure in ≥ 2 loci were selected for re-amplification for all STR loci in order to increase the scoring percentage in the data set. These individuals represented the majority of, but not all the individuals displaying PCR failure at ≥ 2 loci. As a result of re-analysing these 87 individuals, in addition to analysis of the second DNA isolate for 48 of the 50 escapees, > 1000 genotypes in 135 fish were independently scored on two occasions. Of these, no genotyping inconsistencies were observed between original and re-analysis. The resultant overall genotyping success in the STR data set (n = 499 × 15 markers) was > 99% (Additional file 1).

Within the SNP data set consisting of 388 loci × 498 individuals, amplification of individual loci was highly variable, ranging from 0-100% scoring (Additional file 1). Loci displaying less than 95% amplification in the data set were excluded (n = 79). Of the remaining 309 loci analysed for 498 fish, overall genotyping success was > 98%, whereas genotyping success ranged from 66%-100% for loci across individual fish. Nine of the 309 loci were monomorphic in all samples. These loci were excluded from all further analyses, leaving a complete SNP data set of 300 polymorphic loci.

Within the SNP data set, 104 out of 2706 tests of HWE were significant at α = 0.05 (= 3.8%). A total of 294 tests were not computed due to some loci being monomorphic in some samples. Following application of Bonferroni correction in a conservative manner (α = 0.05/300 loci = 0.00017), only 6 of the observed deviations remained significant. These were observed in (locus:sample) 74:A, 202:E, 202:F, 202:G, 207:D, 300:C. Within the STR data set, 16 out of 140 tests of HWE were significant at α = 0.05 (11.4%). These deviations were spread among loci and samples, with all being implicated in a minimum of one significant deviation except loci Ssa289, SsaF43, SSspG7, and samples E, F. Following application of Bonferroni correction in a conservative manner (α = 0.05/14 = 0.0036), none of these deviations remained significant.

A summary of the allelic variation and expected average heterozygosity per sample are presented (Table 1). In total, 600 alleles were observed at the 300 SNP loci whilst 179 alleles were observed at the 14 STR loci. The percentage of the total number of alleles observed within specific samples varied between 92-98% and 53-74% for the SNP and STR data sets respectively. Despite large differences in absolute numbers of alleles between marker sets, corroboration between allelic variation for individual samples (relative to other samples) was observed between marker classes. Expected heterozygosity averaged over loci and samples varied greatly between the 300 SNP data set (He = 0.32) and the 14 STR loci (He = 0.78), although expected heterozygosity, relative to the other samples was similar between marker classes.

Global F_ST values per locus ranged from 0.033-0.115 among the 14 STR loci, and -0.002-0.316 among the 300 SNP loci (Figure 1). 87 of the 300 SNP loci (= 29%) displayed global F_ST values over 0.1 whereas two STR loci (SsaD157 and SSsp2210) exceeded a global F_ST of 0.1 (= 14%). Despite the considerable differences in F_ST among the SNP loci, an analysis using the Bayesian simulation-based test by Foll and Gaggiotti 49 only identified a single SNP as an outlier (Bayesian p < 0.01; data not shown). Hence, there was limited evidence to suggest that the loci might be under diversifying selection in the analysed set of samples.

Within each linkage group, global F_ST per SNP locus varied greatly (Additional file 2). For example, global F_ST ranged between 0.013-0.316 per locus on linkage group d03. Eleven of the linkage groups consisted of 10 or more SNP loci. When mean global F_ST per linkage group was compared among them, no significant differences were observed (Kruskal-Wallis non-parametric ANOVA: P = 0.12). Analysis with a parametric ANOVA gave a similar result (P = 0.4). The majority of tightly linked loci (i.e., those located within the same contig and < 0.1 cM distance) displayed very similar global F_ST values to each other, however, this was not universally true. For example, whilst the three loci located at 13.7 cM on linkage group d02 displayed global F_ST values of 0.026-0.034, and the two loci located at 3 cM on linkage group d14 displayed global F_ST values of 0.0002 and 0.0003, the two loci located at 61.8 cM on linkage group d06 displayed highly contrasting global F_ST values of 0.05 and 0.27.

The genetic relationships between the nine samples collected from farms are presented as UPGMA diagrams (Figure 2). The data set consisting of one SNP per linkage group, selected by highest global F_ST, displayed greater among-sample differentiation than other data sets, however, among sample relationships were remarkably similar for all data sets (Figure 2), including those consisting of all 314 markers combined and 195 mapped SNPs (data not presented).

Using Geneclass, the overall accuracy of self-assignment was 65%, 73% and 73% for the data sets consisting of 14 STR, 300 SNPs and 195 mapped SNPs respectively (Figures 3 and 4). In the STR data set, with the exception of selecting loci starting with the least polymorphic first, the various selection methods only gave small differences in increase of assignment with number of loci (Figure 3), and, almost no further gain in assignment was observed past four loci. In the 300 SNP data set, large differences in the cumulative assignment curve were observed between the different selection methods (Figure 4), furthermore, selection of loci from the 195 mapped SNPs gave the highest overall assignment when approximately 100 loci were included in the analysis (80% assignment). Past this number of loci, the assignment accuracy dropped. Comparing the two marker types, the "best" 15 SNPs selected by BELS matched the level of assignment achieved by the best 4 STR loci selected by allelic variation (and BELS).

When self-assignment simulations were conducted with the SNP loci displaying the highest global F_ST per linkage group (n = 28), overall assignment reached 58% which is similar to the value reported for the best 25 SNP loci selected by global F_ST irrespective of linkage group. However, as the SNP loci displaying highest global F_ST values were spread between linkage groups (Additional file 1), these two sets displayed considerable locus overlap.

Addition of 1-4 STR loci increased assignment for data sets starting with 5, 10 and 25 SNPs selected by BELS, however, for the data set starting with 50 SNPs, addition of STR loci lead to a reduction in assignment (Figure 5). When selecting SNP loci based upon global F_ST, addition of 1-4 STR loci increased assignment in data set starting with up to 100 loci, although a drop in overall assignment was observed when starting with 300 SNP loci (data not presented). For all data sets starting with different numbers of STR loci, addition of up to 50 SNP loci increased assignment when selecting loci with BELS (Figure 6), and global F_ST (data not presented).

"Self-assignment" of the 40 individuals removed from the baseline (10 from A, D, F, G) revealed identical results between the programs Geneclass and STRUCTURE for data sets consisting of 28 SNPs (58%), and 195 SNPs (78%). The latter is an important as STRUCTURE used a marker linkage model, taking marker distance into the computations, whereas Geneclass treated the loci as independent. STRUCTURE outperformed Geneclass for self-assignment of these 40 individuals using 14 STR loci (73% contra 65%), and all 300 SNPs (88% contra 80%).

The absolute accuracy of assignment was lower when computed using a distance based calculation 50, however, the trends in assignment when mixing marker classes were very similar to the trends reported above, although no drop in assignment was observed when STR loci were added to the data set starting with 50 SNP loci.

Direct assignment (Table 2) using all variants of the STR and SNP data sets (including all combinations) demonstrated that nearly all of the escapees originated from sample I. Whilst only a low number of loci were required to directly assign most of the escapees to the sample I, simulations of exclusion from each sample rejecting at P = 0.01 indicated that more loci were required for exclusion of the alternative samples, and, this trend was evident for both marker classes and marker selection criteria (data not presented).

For the data sets consisting of 14 STR, 28 SNP, 195 SNP and 300 SNP loci, k was estimated at 4 or 5 (data not presented), and consequently, assignment of the individuals is presented for k = 3-5 (Additional file 3). The inter-sample relationships revealed by STRUCTURE 2.2 displayed concordance with the UPGMA diagrams for these data sets (Figure 2), furthermore, with minor differences, all four data sets examined displayed a similar pattern of relationships between samples, for each k. These analyses clearly linked the escapees (RF) and sample I into a single cluster separate from all other samples, confirming the assignment results conducted above.