Comparing genomic signatures of domestication in two Atlantic salmon (Salmo salar L.) populations with different geographical origins

Abstract Selective breeding and genetic improvement have left detectable signatures on the genomes of domestic species. The elucidation of such signatures is fundamental for detecting genomic regions of biological relevance to domestication and improving management practices. In aquaculture, domestication was carried out independently in different locations worldwide, which provides opportunities to study the parallel effects of domestication on the genome of individuals that have been selected for similar traits. In this study, we aimed to detect potential genomic signatures of domestication in two independent pairs of wild/domesticated Atlantic salmon populations of Canadian and Scottish origins, respectively. Putative genomic regions under divergent selection were investigated using a 200K SNP array by combining three different statistical methods based either on allele frequencies (LFMM, Bayescan) or haplotype differentiation (Rsb). We identified 337 and 270 SNPs potentially under divergent selection in wild and hatchery populations of Canadian and Scottish origins, respectively. We observed little overlap between results obtained from different statistical methods, highlighting the need to test complementary approaches for detecting a broad range of genomic footprints of selection. The vast majority of the outliers detected were population‐specific but we found four candidate genes that were shared between the populations. We propose that these candidate genes may play a role in the parallel process of domestication. Overall, our results suggest that genetic drift may have override the effect of artificial selection and/or point toward a different genetic basis underlying the expression of similar traits in different domesticated strains. Finally, it is likely that domestication may predominantly target polygenic traits (e.g., growth) such that its genomic impact might be more difficult to detect with methods assuming selective sweeps.


| INTRODUC TI ON
Domestication, "the process by which captive animals adapted to man and the environment he provides," has led to several genetic changes over generations in various animal and plant species (Price, 1984). This evolutionary process may induce similar phenotypic changes in populations of independent origins within the same species when the same phenotypic traits are subjected to similar selective pressures. Phenotypes of these animals have progressively evolved due to the combined influence of domestication selection through reproduction in captivity and human directional selective breeding (Andersson, 2012). This may result in a remarkable phenotypic diversity within domestic species as well as a wide variety of genetic adaptations to both environmental conditions and production systems (Andersson, 2012(Andersson, , 2013. Domestication has thus shaped the genetic diversity of these species throughout history, and their present genomes may contain traceable signatures of selection (Utsunomiya, Pérez, O'Brien, Sonstegard, & Garcia, 2015).
In most fishes, domestication is a recent process compared with any other livestock species (Gjedrem, 2005;López, Neira, & Yáñez, 2015). Therefore, important knowledge gaps still remain about the consequences of domestication on behavior, physiology, and morphology of fishes (Teletchea & Fontaine, 2014) compared to domesticated mammals and birds, for which more data are available (Kelley, Brown, Therkildsen, & Foote, 2016;Lorenzen, Beveridge, & Mangel, 2012). Nevertheless, there is no reason to consider domestication of terrestrial and aquatic animals distinctively (Balon, 2004;Teletchea & Fontaine, 2014). Hence, we might expect that, similar to birds and mammals, selection for the phenotypes contributing to domestication goals has also impacted the extent and distribution of variability within the genomes of fishes. Indeed, domestication has greatly impacted the phenotypes of domesticated aquatic species Detecting genomic signatures of selection is a major goal of modern population genetics (Fariello et al., 2014;Nielsen, Hellmann, Hubisz, Bustamante, & Clark, 2007) as it enhances our knowledge of the molecular mechanisms shaping the genome as well as providing functional information on specific genes/genomic regions that would be of interest for breeding programs (Qanbari & Simianer, 2014). Recently, the progress of high-throughput and cost-effective genotyping techniques has offered a unique opportunity to analyze large datasets of domesticated species to study genome changes in response to domestication events (Druet, Pérez-Pardal, Charlier, & Gautier, 2013;Ma et al., 2015). In livestock species, genomewide analyses have already shown promising results in mapping traits of economic interest, such as genomic regions related to milk production in cattle (Bos tourus) (Qanbari et al., 2011), muscle development in pig (Sus scrofa) (Amaral et al., 2011), coat pigmentation and skeletal morphology in sheep (Ovis aries) (Kijas et al., 2012), or gait and size in horses (Equus caballus) (Petersen et al., 2013). The practical aspect of these population genomics studies lies on the possibility of detecting selected genes associated with traits of economic interest and acts as complement to gene mapping approaches (e.g., genomewide association studies [GWAS]) that may help to further genetically improve these traits on domestic species (Qanbari & Simianer, 2014).
Furthermore, this knowledge is important from an evolutionary perspective by highlighting traits that have been exposed to natural and artificial selection as well as using this information to design and/ or update breeding programs for conservation purposes worldwide (Brito et al., 2017;Cesconeto et al., 2017;Zhao, McParland, Kearney, Lixin, & Berry, 2015).
Recently, methods such as whole genome sequencing or genomewide SNP arrays have enabled the screening of a large part of the genome to detect signatures of selection in domestic and natural populations (Druet et al., 2013;Fuentes-Pardo & Ruzzante, 2017;Ma et al., 2015). Moreover, new genome scan approaches have been developed with the goal of efficiently and accurately identifying genomic footprints of selection out of the thousands of markers screened (Jensen, Foll, & Bernatchez, 2016). Several analytical methods are now available, ranging from population detecting a broad range of genomic footprints of selection. The vast majority of the outliers detected were population-specific but we found four candidate genes that were shared between the populations. We propose that these candidate genes may play a role in the parallel process of domestication. Overall, our results suggest that genetic drift may have override the effect of artificial selection and/or point toward a different genetic basis underlying the expression of similar traits in different domesticated strains. Finally, it is likely that domestication may predominantly target polygenic traits (e.g., growth) such that its genomic impact might be more difficult to detect with methods assuming selective sweeps.

| 139
LÓPEZ Et aL. differentiation analyses, based on F ST calculation (Porto-Neto, Lee, Lee, & Gondro, 2013), to environmental association methods (Cesconeto et al., 2017;Frichot, Schoville, Bouchard, & François, 2013), which aim to detect genetic variants associated with specific environmental factors. Population differentiation methods are expected to detect only strong signatures of selection (e.g., hard sweep), whereas alternative methods such as environmental association would be able to detect subtle signal of selection (e.g., soft sweep), which makes the combination of the two a particularly promising tool for uncovering any genomic footprints of selection (Benestan et al., 2016;Rellstab, Gugerli, Eckert, Hancock, & Holderegger, 2015).
Across the entire Atlantic salmon natural range, European and North American populations exhibit a deep genetic divergence and belong to two divergent glacial lineages which likely separated more than 1,000,000 years ago (Rougemont & Bernatchez, 2018). While these two lineages did not share the same demographic history, domestic populations derived from them have been selected for similar economically important phenotypic traits such as growth, age at sexual maturity, disease resistance, and flesh quality (Gjedrem, 2012), which suggests that domestication may involve parallel evolution. Finding parallel genetic differences between these two populations would reinforce the hypothesis that candidate genomic regions may be linked to molecular processes underlying domestication. Indeed, signals of selection detected among farmed fish belonging to different origins are likely to result from selection rather than having been inherited by chance or from migration of selected allele across breeding strains, which limits false-positive detection.
In Atlantic salmon, previous transcriptomic studies showed that European and North American populations exhibit parallel differential gene expression profiles between hatchery strains and wild populations in the same set of genes (Roberge et al., 2006;Sauvage et al., 2010). Yet, Vasemägi et al. (2012) using 261 SNPs and 70 microsatellites loci found little support for parallel evolution as they found none of the candidate genomic regions potentially associated with domesticated and improved traits overlapped across three different lineages (Ireland, Sweden, and Canada). Using a genomewide analysis with a 6.5K SNP array, Mäkinen et al. (2014) revealed that candidate loci associated with domestication were located on different chromosomes for the North American and European lineages. In this study, we aimed at further investigating the genomic footprints of domestication as well as parallel genomic changes between two pairs of wild/domestic populations of Atlantic salmon with European and North American origins using a 200K SNP chip that may offer higher resolution than achieved in previous studies.
We applied three complementary statistical approaches to identify genomic regions putatively under selection: (a) a genome scan based on population differentiation indices (F ST ) (BAYESCAN), (b) genomewide environmental association (LFMM), and (c) haplotype extended patterns (Rsb). From the set of candidate genes detected by these approaches, we identified their biological function to delineate their potential importance in the process of domestication of Atlantic salmon populations.

| Studied populations
This study was conducted using two wild/domestic population pairs of Atlantic salmon from two different geographical origins: Canada and Scotland, for a total of 183 individuals (see Table 1 for details).
The purpose of this design was to perform two independent comparisons between domestic strains and their wild counterpart populations. Samples from Canadian domestic population (Can-D) were collected from a domestic strain of Atlantic salmon cultured in Chile.
This strain was established from eggs obtained from the Gaspé Bay (where the Saint-Jean River flows), Québec, in the 1950s (Withler, Supernault, & Miller, 2005). Using 11 microsatellites Withler et al. (2005) showed no evidence of introgression from other strains. We presume that fish of this strain were brought to the United States in the late 80s and early 90s and kept in a hatchery located in the state of Washington for about two generations before being introduced to Chile between 1996 and 1998 (J. P. Lhorente, personal communication). This strain has been subjected to intensive selection for rapid growth rate as the main objective of breeding, and also for low incidence of early sexual maturation (J. P. Lhorente, personal  (Dionne et al., 2008). The Saint-Jean River flows into Gaspe Bay, along with York and Dartmouth rivers, and salmon population from Saint-Jean River is the largest one of the three. Furthermore, it has been shown that genetic differentiation among individuals from these three rivers is weak (θ ST (Weir & Cockerham, 1984) = 0.011) (Dionne et al., 2008), enough to consider our samples as a good representation of the possible wild ancestors of domestic Canadian population used in this study. On the other hand, the Scottish farmed population (Sct-D) is a strain cultivated in Chile and it was established with fish from Loch Lochy, located on the West Coast of Scotland. This strain was maintained in a commercial genetic improvement program based in Scotland  and is characterized by a high proportion of early maturing fish and faster growth rate than strains with high proportion of late maturing fish . In 1986, this Scottish farmed population was transferred to Los Lagos Region (42°S 72°O, Chile), presenting a high incidence of fish with early sexual maturity and rapid growth as well, based on field records from 2004. Since then, this population has been adapted to environmental, geographical, and climatic conditions in Chile and selected for rapid growth rate using mass selection for at least six generations (C. Soto, personal communication). The wild population from Scotland (Sct-W) comprised juvenile individuals from South Esk River on the East Coast of Scotland collected in 2011, which is a region with sparse aquaculture activity; therefore, there is a low probability that escaped farmed fish had affected the genetic diversity of wild individuals. We used samples from this location, instead of fish from the West Coast of Scotland, as the latter has an important aquaculture activity and has also been stocked from different sources. Therefore, wild fish from this region does not represent a pure Scottish wild Atlantic salmon. Individuals were collected using electrofishing followed by anesthesia and partial fin-clipping. All fish were released back to the capture location . Fish collecting work has been reviewed both by Marine Scotland Review Committee and the United Kingdom Home Office (Project License 60/4251).

| Genotyping and quality control
Genomic DNA was extracted from fin clips using a DNeasy Blood & Tissue Kit (QIAGEN). For genotyping, we used an Affymetrix Axiom ® myDesign Custom Array of 200K SNPs, which was developed using thirteen European fish (Scottish and Norwegian origin) and seven North American fish, as described by Yáñez et al. (2016).
SNP quality control (QC) was carried out using Axiom Genotyping Console (GTC, Affymetrix) and SNPolisher (an R package developed by Affymetrix), based on SNP clustering metrics and call rate status. Thus, the QC was performed by (a) removing SNP that did not match with high quality clustering patterns, which are defined according to the best practices recommended by Affymetrix through using SNPs with (i) PolyHighResolution (good cluster resolutions for homozygote and heterozygote samples and at least two occurrences of minor allele) and (ii) NoMinorHom (two distinctive cluster with nominor homozygous genotypes), (b) removing SNP loci with call rate lower than 0.95, and (c) discarding individuals with call rate lower than 0.90. Further filtering steps included removing loci deviating from Hardy-Weinberg equilibrium calculated for each population separately, and subsequently, we removed SNPs shared among all populations, and this stage was part of the validation of the SNP chip as described by Yáñez et al. (2016). Loci with minor allele frequency (MAF) lower than 5% within each pair of populations joined (i.e., Can-W and Can-D comprised one group; Sct-W and Sct-D comprised another group) using PLINK v1.07 (Purcell et al., 2007).

| Basic statistics and structure analysis
To investigate population differentiation, we calculated the pairwise Weir and Cockerham's F ST (1984) estimator across all loci among populations, using VCFTools (Danecek et al., 2011). Observed and expected heterozygosity (H O and H E ) were estimated using PLINK v1.07 (Purcell et al., 2007), and confidence intervals for these statistics were calculated using the R package boot with 1,000 bootstrap replicates (Canty & Ripley, 2016). Genomic distribution of F ST was further analyzed by a kernel-smoothing approach using the R package Lokern. A local bandwidth of ~400 was used for fitting the kernel-smoothed regression line. Two clustering methods were used TA B L E 1 Atlantic salmon populations analyzed in this study with geographical origin and sample size (n) to assess the extent of genetic structure among the populations.
First, we performed a principal component analysis (PCA) implemented in the R package adegenet (Jombart, 2008). Second, we used the maximum likelihood estimation of individual ancestries through ADMIXTURE (Alexander, Novembre, & Lange, 2009) software.
ADMIXTURE analysis was run using 2,000 bootstraps, and the number of ancestral populations was set from 1 to 10 (K). The optimal K was selected based on the lowest cross-validation error and a visual inspection of the co-ancestry values for each individual.

| Identification of selection signatures
Three methods were implemented to detect signatures of selection: (a) population differentiation, (b) environmental association, and (c) haplotype extended patterns. First, we used the population differentiation approach with a Bayesian likelihood method, implemented in BAYESCAN v.2.1, which identifies candidate loci under selection using differences in allele frequencies between popula-

tions. The algorithm uses a reversible-jump Markov Chain Monte
Carlo to explore models with or without selection. To estimate the probability that a locus is under selection, the program uses a Bayes factor for two models: one assuming selection and another assuming neutrality given the data (Foll & Gaggiotti, 2008 A Bayes factor between 10 and 32 (log 10 = 1-1.5) for one locus is considered putatively under divergent selection, while between 32 and 100 (log 10 = 1.5-2) and Bayes factors above 100 (log 10 > 2) are considered putatively under very strong and decisive divergent selection, respectively. We used a Bayes factor of 10 (log 10 = 1.30) as threshold to select candidate loci under selection.
Second, we performed an environmental association approach using latent factor mixed models implemented in the software LFMM (Frichot et al., 2013). This method calculates the correlations between allele frequencies and environmental variables. We defined Domesticated and Wild conditions as an environmental dichotomous variable (0 = wild; 1 = domesticated). This method can efficiently estimate random effects due to population history and isolation-by-distance and can reduce false-positive associations compared to genome scans (Frichot et al., 2013). We applied the "latent factor mixed models" algorithm and calculated the |z| scores for all of the SNPs in both independent comparisons for Canadian and Scottish populations, using 20,000 iterations, 10,000 burn-in iterations in the Gibbs Sampling algorithm and a seed of 1,000.
The number of latent factors, which is the best number of clusters describing population structure of the original data, was chosen according to the results of the clustering analyses (i.e., PCA and ADMIXTURE) and as recommended by Frichot et al. (2013).
Finally, we investigated haplotype extended patterns using the

Standardized log-ratio of integrated EHHS (iES) between pairs of populations
(Rsb) analysis (Tang, Thornton, & Stoneking, 2007). This test is based on the extended haplotype homozygosity (EHH) statistic and contrasts EHH patterns of the same haplotype between populations. The SNPs position on the genome is needed estimating EHH, which was possible thanks to the availability of the chromosomal-level genome assembly for Atlantic salmon (Lien et al., 2016). Rsb is defined as the natural log of the ratio between iES POP1 and iES POP2 , where iES is the integrated EHHS (site-specific EHH) for each SNP within each population. As the Rsb values are normally distributed, a Z-test was applied to identify significant SNPs under selection between wild and domestic strains of both Canadian and Scottish origins. A positive value of Rsb indicate iES POP1 is larger than iES POP2 ; therefore, Pop1 has longer haplotypes than Pop2, and hence, positive Rsb scores suggest that selection occurred in the alternative population (domestic population), whereas the negative scores suggest that selection occurred in the reference population (wild population). One-sided p-values were obtained as −log 10 (1-

| SNP annotation and functional enrichment analysis
Genomic regions harboring each locus putatively under selection, detected by at least one method (BAYESCAN, LFMM or Rsb), were interrogated for genes annotated to the Atlantic salmon genome reference ICSAG_v2 (GenBank: GCA_000233375.4) using SnpEff v4.3 (Cingolani et al., 2012). This functional annotation is a complete representation of the genome containing 37,206 high-confidence protein-coding genes that have been assigned to a putative functional annotation based on homology within the SwissProt database (Lien et al., 2016).
Using the salmon transcripts, we performed a blastx on the ze-

| Genotyping and SNP filtering
After genotyping QC, following the conditions recommended by Affymetrix described in Materials and Methods, 159,099 SNPs were categorized in two good quality metrics (a) poly high resolution (good cluster resolution of homozygote and heterozygote samples and at least two occurrences of minor allele) and (b) no-minor allele homozygous (good cluster resolution with no-minor homozygous genotypes).
Thus, a total of 151,509 SNPs were anchored to a unique location on the last version of Atlantic salmon genome assembly (GenBank: GCA_000233375.4)

| Population differentiation and structure analyses
Using the set of markers with MAF > 5% common to all four popu-  Table 2.

| LFMM method
We considered a |z| scores higher than 10 as the threshold to select SNPs associated with domestication using this approach. This cutoff value indicated significant SNP effect at the level of p-value < 10e-7 after applying a standard Bonferroni correction for α = 0.05. We found 115 outliers in Canadian populations ( Figure 2a) and nine outliers in Scottish populations (Figure 2b). In Canadian populations, these outliers were distributed in all chromosomes, except for Ssa14 chromosome ( Figure 3b). The most representative chromosomes were Ssa01, Ssa03, and Ssa04 with 15, 11, and 14 outlier markers, respectively. In Scottish populations, the nine outliers found were distributed in eight chromosomes with Ssa16 being the only chromosomes harboring two outliers (Figure 3e). Details of these results are shown in Table 3.

| Genomic regions putatively under selection distributed across the chromosomes
Overall, kernel-smoothing F ST results did not reveal any obvious island of divergence in common between wild versus farmed salmon in both part of the Atlantic (i.e., Canada and Scotland; Figure 4).
Accordingly, outliers found in each population were distributed differently along the chromosomes ( Figure 5). For instance, more than 24 outliers were found as belonging to Ssa03 chromosome in

| Genomic footprints of selection: Inconsistencies among methods
We uncovered a total of 337 and 270 SNPs putatively under divergent selection for Canadian and Scottish populations, respectively, according to the three complementary methods (Figure 2c).
Overlaps are summarized in Venn diagrams in Figure 2c. In Canadian populations, only Bayescan and LFMM approaches jointly identified 17 outlier SNPs (Figures 2a and 3), whereas for Scottish population, only Rsb and Bayescan approaches found three outlier SNPs in common (Figures 2b and 3). Among these 17 shared outliers between Bayescan and LFMM in Canadian populations, seven regions were found to belong to seven genes (gdnf; opn4x1b2; svil-like; tbck-like; plxnb2-like and two uncharacterized proteins) in Atlantic salmon genome annotation. In Scottish populations, three shared SNPs belong to two predicted genes (brwd3-like and sh3bgrl3-like). Only one common outlier on chromosome 25 was shared between the two geographical origins (Figure 2c), even when we considered all the 337 and 270 outliers detected. This common outlier (Affx-87919237) is located in the gene coding for ubiquitin-conjugating enzyme E2F putative (ube2f).

| Investigating for parallel signal of selection among the North American and European lineages
Overall, we found a total of 138 and 121 genes among the 337 and 270 outlier sequences detected as being potential targets of selection between wild versus domesticated Atlantic

| D ISCUSS I ON
Understanding how the process of domestication may shape the genome of wild to domesticated animals is particularly useful for enhancing our knowledge on how human-driven selection may induce genetic changes as it may provide a practical framework to guide genetic improvement practices (Wang, Xie, Peng, Irwin, & Zhang, 2014). Here, we performed a large genomewide scan, using a dense SNP array and three outlier methods, to investigate genomic footprints of domestication in two geographically isolated Atlantic salmon have diverged independently for possibly as much as 1 million years (Rougemont & Bernatchez, 2018), it is likely that this had led to different standing genetic variation upon which selection may be acting, limiting the potential for parallel evolution at the genome level in this study system. This is in agreement with previous population genomic work on Atlantic salmon (Mäkinen et al., 2014;Vasemägi et al., 2012) and, more broadly, on several studies showing that selection may act faster on standing variation compared to new mutations (Barrett & Schluter, 2008). Limited evidence for parallel impact of artificial selection at the genome level contrasts with the pronounced pattern of parallelism previously documented at the transcriptome level (Roberge et al., 2006), suggesting the different genomic architecture may result in similar pattern of gene expression as well as parallel phenotypic changes occurring during domestication. Clearly, shedding light on the complete genomic basis of domestication of Atlantic salmon would require using alternative methods rather than genome scan and collecting a larger dataset with known phenotypes. Future studies testing for parallel evolution in Atlantic salmon should consider this avenue.

| Limitations of genome scan methods
Traits of interest selected in aquaculture programs, such as growth and fat content, may be polygenic, that is, controlled by many genes with high variance of effect size (Gagnaire & Gaggiotti, 2016).
Polygenic selection may imply a subtle change of allele frequencies at several loci that would be hard to detect using genome scan methods that mainly focus on moderate to high shift in allele frequency at few independent loci. For instance, Bourret, Dionne, and Bernatchez (2014) genotyped 5,568 SNPs to test for differential allelic and genotypic frequencies between juveniles (smolts) migrating to sea and adults (grilses) returning to freshwater after 1 year at sea. Although numerous outliers were identified by the single-locus analysis, no evidence for parallel, temporally repeated selection was found. In  Information Table S3) were detected in Scottish population, which is a lower number than those detected using filtering options, revealing the suitableness of applying these filters to discard low-quality genotypes.
Alternative models to common genome scan approaches, such as a quantitative genetic framework, provide powerful tools to investigate polygenic selection. Earlier Atlantic salmon studies of quantitative trait loci (QTL) mapping have already reported a large number of different QTLs for growth (Baranski, Moen, & Våge, 2010;Gutierrez et al., 2012;Houston et al., 2009;Reid, Szanto, Glebe, Danzmann, & Ferguson, 2005), although genomewide studies revealed very low levels of association between markers and growth Gutierrez, Yáñez, Fukui, Swift, & Davidson, 2015).
This indicates that growth-related traits are likely to be controlled by a number of population-specific loci of low to moderate effect with an important polygenic component (Tsai et al., 2015), making it difficult to compare our results with those obtained from GWAS or QTL mapping. Future studies combining population genomic and quantitative genetic framework would then provide a more inclusive approach for uncovering the molecular footprints of domestication in Atlantic salmon. Furthermore, the complete functional annotation of the Atlantic salmon genome will undoubtedly enhance opportunities for a comprehensive understanding of the effects of domestication in Atlantic salmon populations (Macqueen et al., 2017).

| Standing genetic variation and loss of genetic diversity
Of the 151,509 SNPs successfully genotyped, 61,199 and 130,586 SNPs were polymorphic for Canadian and Scottish populations, respectively. This striking difference on the number of polymorphic SNPs is in accordance with earlier studies based on various types of markers (mtDNA, microsatellites, AFLPs, SNPs) showing that North American Atlantic salmon populations have lower genetic diversity than European populations Mäkinen et al., 2014). This may also be explained to some extent by ascertainment bias as the 200K SNP Array used in the present study was created based on genomic information mostly from European salmon . Both lineages only shared 55,406 SNPs in common and the genetic diversity analysis of these SNPs revealed that the lowest and highest levels of heterozygosity were found in Canadian domestic and Scottish wild population, respectively. Indeed, lower levels of genetic diversity in both domestic populations compared to their wild homologs were expected as most of domestic populations experience an inevitable loss of genetic diversity due to both selective breeding as well as the absence of genetic connectivity with other populations as found in nature (Baumung, Simianer, & Hoffmann, 2004;Johansson & Rendel, 1968

| Detecting the impacts of domestication with genome scans
Using a genomewide SNP array, we found limited evidence of potential signatures of selection that were jointly detected between domestic and natural populations, in Canada and Scotland, considering the outcomes of three different outlier tests. This lack of congruence among methods is likely to result from the statistics underlying each approach: LFMM is based on linear associations, Bayescan is an F ST differentiation method, and Rsb is a haplotypes comparison method. As LFMM is based on linear model and it detects a few numbers of outliers in our study system, this approach may not be the most appropriate approach for this kind of experimental design. More particularly, Rsb test is likely to detect more recent selection signatures in comparison with F ST -based methods (Oleksyk, Smith, & O'Brien, 2009). On the other hand, Bayescan was also the method that resulted in the highest number of outliers identified but it should be noted that we used a permissive threshold (PO = 10), as used by

| Investigating the existence of parallel genomic footprints of Atlantic salmon domestication
Over the past 50 years, intense artificial selection in Atlantic salmon populations has led to the development of strains specialized in certain phenotypic traits which has caused large phenotypic and genotypic changes between wild and domestic populations (Glover et al., 2009). Despite Atlantic salmon populations having been intensely selected for growth-related traits on both sides of the Atlantic ocean, we found little overlap between the outlier genomic regions identified in previous studies of other Atlantic salmon strains selected for high growth rate Gutierrez, Yáñez, Fukui, et al., 2015). These population-specific signals suggest that selection may have acted upon different genes, which was already shown by one study documenting genomewide footprints of pig domestication (Amaral et al., 2011). More particularly, several population genomic studies of Atlantic salmon populations have suggested that the same phenotype may arise from different genetic pathways among geographically isolated populations (Elmer et al., 2014;Mäkinen et al., 2014;Perrier, Bourret, Kent, & Bernatchez, 2013;Pujolar, Ferchaud, Bekkevold, & Hansen, 2017;Vasemägi et al., 2012). Evolution of complex parallel phenotypes can indeed arise from different evolutionary routes and this is likely to happen when inbreeding and genetic drift play a greater role than selection.

| Parallel genomic regions detected as being putatively under artificial selection
Despite limited evidence for parallelism at the genome level, evidence for parallel evolution was observed at a few potentially important genes. Here, genomic regions harboring collagen alpha-1XIII chain gene (coda1) were identified in both comparisons. There is evidence that collagen genes may be involved in the response of Atlantic salmon to sea lice as initiators of inflammatory cytokine signaling (Castillo-Briceño et al., 2009;Correa et al., 2016;Krasnov, Skugor, Todorcevic, Glover, & Nilsen, 2012). Therefore, this gene might be involved in the immune response to cope with specific diseases present in aquaculture environment and may be a relevant target of selection. Despite none of the farmed populations having been directly selected for traits associated with disease resistance, farmed populations are often subjected to high levels of pathogens, such as parasite infections, which are known to be among the strongest selective forces driving the evolution of host populations (Roberge et al., 2006;Zueva et al., 2014). A second common gene was ubiquitin-conjugating enzyme E2F putative (ube2f), coding for ubiquitin-conjugating enzyme E2-F. Growth and development of skeletal muscle undergo breakdown and replacement of proteins during different periods, and in teleost fish, this process involves E2ubiquitin-conjugating enzymes and E3-ubiquitin ligases (Johnston, Bower, & Macqueen, 2011). Therefore, this gene could be involved in muscle development and growth in Atlantic salmon, which is an important trait selected for enhancing salmon aquaculture production. This gene family was also suggested to be potentially involved in immune response of marine vertebrates (Núñez-Acuña, Aguilar-Espinoza, Chávez-Mardones, & Gallardo-Escárate, 2012), and more particularly, ubiquitin genes have already been detected to be potential targets of selection in the domestication process of shrimp (Rotllant et al., 2015), auroch (Braud et al., 2017) and rice . Finally, the autism susceptibility 2 protein-like (LOC106585083) was also found to overlap between continents.
This gene (Auts2-like) has been associated with autism and mental retardation in humans (Bedogni et al., 2010;Oksenberg & Ahituv, 2013). Interestingly, Auts2 has been found to be under selection in domestic cattle breeds (Consortium, Bovine HapMap, 2009), suggesting that domestication may act on behavioral traits in salmon as in other domestic animals (Clutton-Brock, 1999). Finally, a potential parallel signal of selection was located in the transient receptor potential cation channel subfamily M member 3-like (LOC106580056), which belongs to Trpm3, a gene associated with the reception of noxious temperature stimuli in mammals (Vriens et al., 2011). Transient receptor potential cation channel subfamily genes are involved in melanocyte pigmentation (Cieslak, Reissmann, Hofreiter, & Ludwig, 2011). Therefore, this gene may play a role in the immune system of domesticated Atlantic salmon and Kittilsen, Johansen, Braastad, and Øverli (2012) showed that pigmentation seems to be correlated with the development of the ectoparasitic lice in Atlantic salmon.
Interestingly, Yang, Li, Li, Fan, and Tang (2014) also detected selection signature linked to domestication process in this gene by comparing Chinese indigenous and commercial pig breeds.

| Biological function of nonparallel genomic regions identified as putatively being under artificial selection
Some of the genes that overlap between statistical methods in Canadian populations (Bayescan and LFMM) included supervillin-like gene, which is a protein involved in actin and myosin II assembly that promotes cell growth (Fang & Luna, 2013). This gene has been shown to be involved in muscle fiber type determination in large white pigs (Zhu et al., 2016). We suggest that this gene may be associated with muscle development in Atlantic salmon and be under selection for improving growth-related traits. We also identified plexin-b2-like, which in mammals is involved in several processes during development of the nervous, cardiovascular, renal, and skeletal system (Worzfeld et al., 2014). Another plexin from the same subfamily has been associated with behavior-related traits, such as tameness and temperament in rat and cattle (Friedrich, Brand, & Schwerin, 2015;Heyne et al., 2014). This finding is in line with the identification of the gene autism susceptibility 2 protein-like and both suggest that domestication could be acting on behavior-related traits. On the other hand, We detected a potential gene candidate in Scottish populations, brwd3, that has been associated to mental retardation in humans (Field et al., 2007) and which was also identified as a target of selection during the domestication process in cattle (Consortium, Bovine HapMap, 2009). Other genes detected in this study included myopalladin-like gene in Scottish populations, which has been associated with late maturation in Atlantic salmon (Gutierrez et al., 2014). The longchain fatty acid-CoA ligase 4-like gene, located in Ssa04, which belongs to Acsl family of genes, was found as a potential gene influenced by domestication in Scottish populations. This gene is involved in lipid metabolism, although the activation of long-chain fatty acids for synthesis and degradations of cellular lipids (Golej et al., 2011). In rainbow trout, lipid metabolism has been reported to be associated with growth (Xu et al., 2011). Similarly, another gene from the same family, the longchain fatty acid-CoA ligase 1 has been shown to be associated with growth in clam (Meretrix meretrix, Dai, Huan, Wang, Xia, & Liu, 2015).
No other overlaps were found with previous studies related to sexual maturity in Atlantic salmon, for both Canadian and Scottish populations. Nevertheless, we found putative outliers in the bromodomain-containing two genes (br2) and the zona pellucida (ZP) sperm-binding protein 3-like gene (Zp3) in Canadian and Scottish populations, respectively, both localized in chromosome Ssa27. The bromodomain-containing 2 gene may be involved in spermatogenesis or folliculogenesis (Rhee, Brunori, Besset, Trousdale, & Wolgemuth, 1998), while ZPs are responsible for the initial sperm binding and the subsequent induction of the acrosome reaction that allows sperm penetration in mammals (Lin, Roy, Yan, Burns, & Matzuk, 2007). Thus, these genes are interesting functional candidates for reproduction-related traits under selection on these populations. To better understand the molecular functions of these genes, we investigated their GO classification. Many genes were categorized into cellular process and metabolic process. Both categories can involve cell growth or anabolic/catabolic process resulting in cell growth, which suggest that these genes could be associated with growth in Atlantic salmon. Admittedly, these observations need to be explored and verified in further studies.

CO N FLI C T O F I NTE R E S T
None declared.

DATA A RCH I V I N G
Data available from the Dryad Digital Repository: https://doi. org/10.5061/dryad.60b9p56.