Aquaculture is currently one of the fastest-growing agricultural production sectors worldwide. FAO (2019) reported that half of the fish used for human consumption were produced in the aquaculture sector, and in 2025, the expected total fish production will be approximately 200 million tons (OECD/Food and Agriculture Organization of the United Nations 2014). Among the 250 farmed fish species around the world, only 10% have been subjected to selection programmes (Teletchea and Fontaine 2014; FAO 2019). Worldwide, the current number of breeding programs for different species is approximately 60 (Abdelrahman et al. 2017; Houston 2017), sharing over 80% of the total production (FAO 2019, FAO Fisheries and Aquaculture Department 2019).

Initially, in most selection programmes, growth rate was the most-selected trait because of its generally medium heritability and ease of application. The genetic improvement rate per generation was generally 4–5 times higher than those obtained in other domestic animals (Gjerde and Korsvoll 1999; Thorgaard et al. 2002; Kause et al. 2005; Georges et al. 2019). While growth rate is one of the most economically important traits, other characteristics, such as feed conversion efficiency, carcass-quality characteristics and disease resistance, also have been gradually incorporated into selection programmes (Gjedrem et al. 2012). These traits are more difficult to select for because of lower heritability, post-mortem measurement and need for use of selection schemes based on sib selection (Yáñez et al. 2015) (which uses only the genetic variance between families) (Hill 2013). Except for a few species, such as Atlantic salmon, rainbow trout and Nile tilapia (Ponzoni et al. 2011), the availability of selected lines and the development of breeding programmes in the aquaculture sector are still in their infancy (Houston 2017). The situation is more favourable in Europe, where over 80% of the production of six main species (Atlantic salmon, rainbow trout, European seabass, gilthead seabream, carp (common carp (Cyprinus carpio) and silver carp (Hypophthalmichthys molitrix)) and turbot (Psetta maxima) derive from improved stocks (Janssen et al. 2018). There is enormous potential to improve worldwide fish production further using selection techniques. Despite the considerable diversity of production volumes, the genetic variability within species and high fish fertility allow the use of flexible breeding plans without the need for expensive multiplication centres (Houston 2017).

Quantitative trait loci (QTLs) of the main productive traits, sexual maturation, fillet characteristics, and disease resistance can be used in genetic improvement programmes through the genetic marker-assisted selection (MAS) approach. Favourable alleles can be identified using SNP arrays, Genome-Wide Association Study (GWAS) and linkage association mapping methods (Goddard and Hayes 2009). The availability of new-omics methods, such as last-generation genome sequencing techniques (e.g. long-range technologies (Pacific Biosciences and Oxford Nanopore) (Lien et al. 2016; Christensen et al. 2018a) and linked reads (Takayama et al. 2021), 10X microfluidic platform (Wang et al. 2020; Zheng et al. 2016), new genome assembly methods (long-range scaffolding, Hi-C (high-throughput chromatin conformation capture, Dovetail and optical mapping, Bionano Genomics) (Burton et al. 2013; Howe and Wood 2015; Putnam 2016; Mantere et al. 2019), transcriptome and epigenetic analysis (RNA-seq, ChIP-Seq), new genotyping techniques (high- and low-density SNP arrays) (Palti et al. 2015a) and genotyping-by-sequencing techniques (GBS, RAD-Seq), can revolutionise the aquaculture breeding sector. These new methods have the potential to identify the most important genetic variants (SNPs) that influence phenotypic expression of the main traits (Goddard and Hayes 2009; Ellegren 2014). Genetic markers can be used for various purposes, such as the analysis of the genetic variability of QTLs (Palti et al. 2015b), MAS, GWAS and genomic selection (Eggen 2012; Liu et al. 2019). The first application of MAS was carried out on Atlantic salmon for the selection of the QTL for resistance to infectious pancreatic necrosis (IPNV) (Houston et al. 2008; Moen et al. 2009). In Table 1 are reported the main definitions and abbreviations used in the text.

Table 1 List of the main definitions and abbreviations used in the text

Continuous improvement of selective breeding

The first modern breeding programme that used family pedigree information was applied to Atlantic salmon (Salmo salar) and started in 1972 in Norway (Gjerde 1986, Gjedrem 2000; Bentsen and Olesen 2002). Initially, the programme focused only on improvement of growth rate or standard growth rate (SGR), but after obtaining positive results, late sexual maturation, carcass quality and resistance to various diseases were also included in the selection index. This programme suggested that growth rates can be improved by 10-20% per generation in most fish species (Olesen et al. 2015; Gjedrem and Rye 2018). This result can be achieved through a simple mass selection programme that does not use family information (Gjedrem et al. 2012). Such programmes have the advantage of being much less expensive, but several traits cannot be evaluated on live candidates, such as slaughter yield and disease resistance. Furthermore, the use of family information allows a more precise estimate of the genetic response at the farm level and faster progress (Sonesson et al. 2010; Vela-Avitúa et al. 2015). In various breeding programmes of Atlantic salmon and rainbow trout, 100-200 families were reared in separate tanks up to the size at which fish can be tagged (Fernández et al. 2014). In this case, the selection process has limited accuracy, as it is based on the choice of the estimated breeding values (EBVs) of the families (use of only 50% of the genetic variability and selection of related individuals) (Ødegård et al. 2011, 2014; Nirea et al. 2012).

At the end of the 1990s, a radical change in fish breeding methods was made by the introduction of a new technique for the identification of parents by means of molecular markers. This method uses the genetic markers of offspring and parents (typically 10-15 microsatellite markers or 100 SNPs). After the statistical analysis, it is possible to assign all the offspring to their parents and thus to keep track of the pedigree (Vandeputte and Haffray 2014). Today, in Europe, in relation to the six main farmed fish species, among 36 selection programmes, 12 use parental assignment as a tool for tracing the pedigree (Janssen et al. 2017). To date, however positive results using marker-assisted selection (MAS) have been obtained only for simple traits in which the main QTL is known (Jørgen et al. 2011; Yue 2014).

The mating system used in a breeding programme is a key factor because it can influence the inbreeding rate (Sonesson et al. 2012), and the selection response (Hely et al. 2013; Sonesson & Ødegård 2016). D'Agaro et al. (2007) compared different mating systems using the optimum contribution selection method (Sánchez-Molano et al. 2016; D’Agaro 2017). The results of the simulation showed that the factorial mating system allows us to obtain the lowest inbreeding rate and the highest selection response. Skaarud et al. (2014) showed that increasing the number of families from 50 to 400 and using the optimum contribution selection method increases the response to selection. Sonesson and Ødegård (2016) indicate that if the number of families increases from 100 to 1000 using 5000 SNPs, the inbreeding rate of the population decreases from 0.014 to 0.006, and the genetic gain (ΔG) is in the range 0.17–0.40. The larger number of families led to higher genetic gains because of the increased selection intensity, while the 5,000 SNPs allowed accurate prediction values to be obtained. From recent studies, it appears that 200 families are the optimal number to be used in fish breeding schemes (Skaarud et al. 2014). Skaarud et al. (2011) suggested using 8–200 offspring per family. The main traits selected in the fish breeding programmes are growth rate, morphology, disease resistance, slaughter yield and product quality; among these, substantial advances can be achieved in growth (+10–20% per generation) and disease resistance (+12.5% survival on a specific disease per generation) (Gjedrem 2000, 2012).

Genomic resources

Next-generation sequencing (NGS) refers to the set of DNA sequencing technologies that have in common the ability to sequence in parallel millions of DNA fragments. The whole process includes several steps: DNA extraction, sequencing and bio-informatic analysis.

The magnetic beads method is the most commonly used technique for DNA extraction for NGS analysis (McGaughey et al. 2018). This method greatly simplifies the extraction procedure and improves the size selection process. The choice of sequencing method depends on several factors, such as the desired coverage depth, error rate, simultaneous multiplexing and throughput capacity.

NGS technologies, also known as high-throughput sequencing (HTS) technologies, have evolved rapidly in the last two decades, offering a greater number of sequenced bases at a lower cost (Guppy et al. 2018; Robledo et al. 2018a). In 2006, second-generation NGS technology was started, which is able to perform millions of polymerisation reactions in parallel (Metzker 2010). The main platforms now available are MiSeq, HiSeq and NovaSeq from Illumina (San Diego, USA). This type of technology is based on the fragmentation of DNA and amplification in parallel with multiple reactions, obtaining short reads between 100 and 300 bp. The distance between the pairs of reads is called insert size (mate paired: 2–5 kb and paired end < 1 kb). In 2013, third-generation NGS techniques, also known as single molecule sequencing (SMS), started. SMS produces longer reads (5–50 kb), with higher costs (>2000 US$ per Gb) and higher error rates. With this method, a single DNA molecule (without enzymatic process or hybridisation) is sequenced, and no library amplification is required. The main platforms are Pacific Biosciences (PacBio) (Lien et al. 2016; Christensen et al. 2018a) and Oxford Nanopore Technologies (ONT) (Jain et al. 2016). These techniques are used mainly to complete the genomic sequence (gap filling) (since repetitive and tetrasomic regions are difficult to assemble, especially for fish) and for scaffolding sequences previously assembled using second-generation technologies (Xiao and Zhou 2020). The difficulties in assembling fish genomes derive from the recent genome duplication process in certain lineages, notably salmonids, the presence of high levels of heterozygous loci and long repetitive elements. The length of the assembled reads should be greater than the size of the haplotype blocks (ranging from 20 to 200 kb). The PacBio platform now routinely generates reads with an N50 >1 Mb. This method has recently reduced the error rate by means of a new sequencing technique called “circular consensus” and the production of high-fidelity reads of 15 kb. The software packages generally used for the bioinformatic analysis of third-generation sequencing are Canu (Koren et al. 2017) and Mecat (Xiao et al. 2017). The results obtained are subsequently cleaned and analysed with other software packages such as Racon (Vaser et al. 2017), Nanopolish (Loman et al. 2015), Pilon (Walker et al. 2014) and Busco (Simao et al. 2015). In 2016, a new third-generation mapping technology was used by Dovetail Genomics (Santa Cruz, CA, USA). This technique allows us to obtain a high-quality scaffolding process at the chromosomal level using two types of libraries (Chicago® and Dovetail™ Hi-C) generated from the chromatin proximity ligation process (Lieberman-Aiden et al. 2009; Putnam 2016). Through these new methodologies, it is possible to reconstruct long haplotypes (phasing), identify structural variations (SVs) (Edge et al. 2017) and distinguish chromosome sets (Weisenfeld et al. 2017).

New sequencing techniques have allowed the assembly of several reference genomes for salmonids (Gonzalez-Pena et al. 2016; Alhakami et al. 2017): Atlantic salmon (Gilbey et al. 2004; Davidson et al. 2010; Lien et al. 2016), rainbow trout (Berthelot et al. 2014), Chinook salmon (Oncorhynchus tshawytscha) (Christensen et al. 2018a), coho salmon (Oncorhynchus kisutch) and Artic charr (Salvelinus alpinus) (Christensen et al. 2018b). Currently, several high-density genomic maps are available for the main salmonid species: Atlantic salmon (Moen et al. 2008; Phillips et al. 2009; Lien et al. 2011; Gonen et al. 2014; Gonen et al. 2015a; Tsai et al. 2016b; Yáñez et al. 2016); coho salmon (McClelland and Naish 2008); Chinook salmon (Brieuc et al. 2014; McKinney et al. 2016); rainbow trout (Young et al. 1998; Sakamoto et al. 2000; Nichols et al. 2003; Guyomard et al. 2006, 2012; Rexroad et al. 2008; Palti et al. 2012; Silva et al. 2019); brown trout (Salmo trutta) (Gharbi et al. 2006; Leitwein et al. 2017); brook charr (Salvelinus fontinalis) (Sutherland et al. 2016) and Arctic charr (Salvelinus alpinus) (Woram et al. 2004). Several important initiatives were established, such as the GRASP (Consortium for the Genomic Research of all Salmonids) initiative (MacQueen et al. 2017) and the Salmobase platform (Samy et al. 2017). These initiatives have allowed the creation of EST (Expressed Sequence Tag) databases, gene expression platforms (microarrays) and SNP array chips. The recent publication of the PhyloFish database (Fouchécourt et al. 2019) represents another important contribution for 15 fish species. A reference genome sequence is essential for understanding all the complex characteristics in cultured and natural fish populations (Gonzalez-Pena et al. 2016).

The advent of NGS technologies has facilitated the identification of a large number of DNA markers randomly distributed throughout the genome, making the use of multilocus data (multiple concatenated loci) more suitable. Numerous SNPs have recently been identified for different samonid species (Adzhubei et al. 2007; Sanchez et al. 2009; Hohenlohe et al. 2011; Houston et al. 2014, Houston 2017; Salem et al. 2012). The combined use of whole-genome sequencing and the analysis of the transcriptome (RNA-seq) (Sutherland et al. 2014; Shi et al. 2019) allows us to obtain a very large number of SNPs (He et al. 2017; Robertson et al. 2017). Note that if the goal is the identification of SNPs in non-model species, RAD-tag analysis is usually sufficient. SNPs can be analysed with SNP arrays or RAD-tag methods. The best method to genotype large numbers of SNPs is the design of SNP high-density (HD) arrays (Hoggart et al. 2008; Matukumalli et al. 2009; Ramos et al. 2009; Cheng et al. 2015).

In recent years, several basic research and applied projects on Atlantic salmon and rainbow trout were carried out. They include the development of an SNP HD array for Atlantic salmon (Houston et al. 2014), comparative whole genome analysis (WGA) of similar species (Gutierrez et al. 2018), resequencing of the rainbow trout genome (Oncorhynchus mykiss) (Gao et al. 2018) and de novo transcriptome analysis of Atlantic salmon, brown trout, and Arctic charr (Salvelinus alpinus) (Carruthers et al. 2018). Applications of SNP arrays in salmonids have been reported in several studies concerning population diversity, QTL mapping and genomic selection (Balding 2006; MacLeod et al. 2010). In the last 10 years, several SNP arrays with different marker densities and formats have become commercially available: a 130K SNP array for rainbow trout, a SNP HD array for Atlantic salmon (Houston e al. 2014; Yáñez et al. 2014) and additional SNP chips for other salmonid species (Lien et al. 2011). SNP arrays are now routinely used for genotyping rainbow trout populations and Atlantic salmon (Houston et al. 2014; Yáñez et al. 2014). Zhang et al. 2018a, 2018b) used a 57K SNP array to genotype several salmonid species. A substantial reduction in genotyping costs can be achieved by means of inference methods such as imputation. Imputation techniques allow the transfer of genomic information from an SNP HD array to a low-density chip (Marchini et al. 2007; Habier et al. 2009; Hickey et al. 2012). NGS technologies also can be used to study non-model species since, in this case, it is not necessary to have previous genomic information. In many cases, a complete genome sequence is not necessary, as the genes are organized in linkage blocks, and a high-density genetic map of the species is available. The availability of uniformly distributed genetic markers allows us to obtain the necessary coverage. Furthermore, the low-pass whole-genome sequencing method (LP-WGS) (0.1-4x) can be used as an alternative to genotyping array methods. This less-expensive technique is based on NGS with low coverage and subsequent imputation. Resequencing of specific genes (using NGS techniques) also can be used to study SNPs, INDELs (insertion–deletion mutations), CNVs (Copy Number Variations) and rearrangements in the same species with lower cost and higher coverage.

The most popular genomic methods used in selection breeding programmes are reduced-representation library (RRL) methods. These methods include the preparation of reduced libraries through the use of restriction enzymes (RAD-seq), transcriptome sequencing (RNA-seq), targeted amplicon sequencing (TAS) and hybrid enrichment with the use of probes (Gonen et al. 2015b; Robledo et al. 2017). The technical basis of RRL is the use of restriction enzymes to digest the genome and then to select the digested fragments. The advantage of RRL methods is the reduction of the cost of the analysis with high coverage (Amish et al. 2012). Genotyping errors can be eliminated by means of specific software such as Angsd (Korneliussen et al. 2014) and NgsCovar (Fumagalli et al. 2014). The most popular RRL methods are restriction-site-associated DNA sequencing (RAD-Seq) (Baird et al. 2008; Davey et al. 2013; Cruaud et al. 2014; Hand et al. 2015; Rochette et al. 2019) and genotyping by sequencing (Genotyping By Sequencing - GBS) (Elshire et al. 2011; Dodds et al. 2015; Gorjanc et al. 2015; Brouard et al. 2017). For instance, de novo RAD-Seq analysis is used in non-model species because it does not require a priori knowledge of the reference genome sequence (Gonen et al. 2015b). In addition, nGBS (normalised genotyping by sequencing) involves the use of restriction enzymes for the digestion of DNA, producing a large number of fragments for each sample. These fragments are then sequenced, producing several million reads per sample. The RAD-Seq method was used on Atlantic salmon (Houston et al. 2012) and rainbow trout by Palti et al. (2014) in selection programmes. The GBS method can be used for SNP genotyping (Elshire et al. 2011; Andrews et al. 2016; Li and Wang 2017; Robledo et al. 2017). Compared to SNP arrays, this method requires more rigorous quality control of genomic data. In fact, some spurious or missing data may lead to erroneous conclusions (Gorjanc et al. 2015; Andrews et al. 2016). These errors may occur during sequencing and alignment analyses (Nielsen et al. 2012; Davey et al. 2013; Sims et al. 2014; Verdu et al. 2016). Several studies on fish species have shown these problems in relation to high polymorphism rates and the presence of highly repetitive elements in the genomes (Lal et al. 2016; Yuan et al. 2017). Methods that are currently available for reducing the aberrant data obtained during RAD-Seq and GBS analysis are based on quality control checks (QScore ≥ 30) (Lal et al. 2016), verification of Hardy-Weinberg equilibrium, and verification of the presence of paralogous genes or SNPs with low coverage (Catchen et al. 2013; Bianco et al. 2014; Torkamaneh et al. 2017). Improving the accuracy of genomic data obtained with GBS and RAD-Seq methods is particularly important for fish populations with limited genetic resources. The interest of the GBS and RAD-Seq methods will remain high in the fish breeding sector until the cost of SNP HD arrays becomes more affordable.

Genome-wide association studies

Genome-wide association studies (GWAS) have been proposed as an effective approach for the identification of many causative mutations and genetic factors that constitute the main quantitative traits (Stram 2004; Meuwissen 2007; Witte 2010; Daetwyler et al. 2010). Unlike linkage studies, which consider the phenomenon of inheritance of chromosomal regions linked to the presence of a trait within a family, association studies consider instead the difference between the frequency of SNPs affecting the trait of interest in a population and then the incorporation of the information into the calculation of the breeding value of animals (Tsai et al. 2015). Association studies may be conducted through two approaches: direct and indirect. A direct association study is to catalog and test one by one all the possible causal mutations. However, the direct approach presents some practical problems. This strategy involves genome-wide identification of all genes (up to 30,000 genes) as well as of all SNPs. For these reasons, the use of the direct method is limited to a few cases and has almost always been replaced by the application of the indirect method. The indirect strategy avoids the need to catalog all mutations that could potentially give predisposition to a given trait and instead relies on the association between a giver phenotype and markers located near a strategic locus. These associations are obtained from studies of linkage disequilibrium (LD) between marker loci. The indirect strategy then employs a dense map of polymorphic markers to explore the genome in a systematic way.

The choice of markers differentiates further the indirect approach in two different strategies. In the first, markers are chosen very close to exon regions of known genes. The second employs markers located in large regions, virtually anywhere in the genome, thus considering the chromosomes in their entirety, including intronic regions. The choice of the marker falls on bi-allelic SNPs because of their high frequency in the animal genomes, for the low rate of mutation and for the ease with which it can be analyzed. Linkage means the presence of genes at close loci on the same chromosome.

LD is a combination of alleles at two or more loci that occurs more often than it would happen by chance. The presence of a LD thus indicates co-segregation of two markers. In generally, the LD between two SNPs decreases with the physical distance, and the extent of LD varies strongly among the regions of the genome. LD analysis is a valuable tool for fine mapping. The LD is strongly influenced by the local recombination rate, distance between the gene and markers and is correlated with other factors such as the content of bases G and C, the gene density and the presence of SINE or Alu1 repeats (Smith et al. 2005). A high degree of similarity was observed between several populations (in regions classified as high LD); this was due not only to past historical shared sequences between populations, but also by the fact that the LD may be influenced by the features of the local sequence. It is generally confirmed that while the bottlenecks of the population, geographical sub-divisions and selection tend to increase the extension of the LD in the single population, population growth and random mating tend to decrease the extension of LD in the genome (Daetwyler et al. 2010). In addition to these factors, the extension of the LD in a particular region of the genome can also be influenced by the physical characteristics of the surrounding DNA sequences. Some types of sequences, such as, for example, rich sequences of bases G and C, may be associated with higher rates of mutation, two phenomena that can directly lower the level of surrounding LD. LD is generally weaker near the ends of chromosomes, probably due to the high recombination rate in the neighborhood of these regions and is stronger around the centromeres and in other intermediate portions of each chromosome (large portions of the repeated genome), where recombination rates are on average lower. Furthermore, the LD is generally stronger in the large chromosomes, which have a recombination rate lower and weaker in the small chromosomes, that have a higher recombination. The presence of repeated sequences in a region is strongly associated with LD and this correlation is significantly increased when we consider large windows of genome. The presence of repeats LINE is associated to an increase of the level of LD while other types of repetition, in particular SINE and Alu, are strongly associated with a lower level of LD. To further characterize the relationship between LD and changes in the sequence, Smith et al. (2005) have divided the genome into quartiles according to the estimated levels of LD in each 100-kb window. The lower frequency allele (MAF) of SNPs in each of the four quartiles (corresponding to the regions of LD high, middle and bottom) were very similar. The study found that the content of G and C decreases gradually in correspondence with the increase of the degree of LD (4349 to 10,000 nucleotides of GC bases in the quartile with the low LD and 3904 to 10,000 nucleotides GC bases in the quartile with the highest LD). It was also noted that the genes are significantly more concentrated in the quartile with the highest and lowest quartiles of LD compared to the intermediate level. This suggests that while for some genes may be advantageous to be located in regions of strong LD, where no recombination takes place, other genes may be favored by the diversity of haplotypes. Smith et al. (2005) have also provided a functional annotation of each gene, depending on of their location in the quartiles with the highest and lowest LD, using the database Gene Ontology. Genes associated with immune response (including genes involved in the inflammatory response and in response to pathogens and parasites) and neurophysiological processes (including sensory perception) are often located in regions with a low degree of LD. In contrast, genes associated with response to DNA damage, cell cycle, or to the metabolism of DNA and RNA appear more often located in regions with high LD. In fact, these genes represent conserved biological processes where recombination and mutation become deleterious changes that are generally removed through natural selection. Despite the fact that LD constitutes a powerful tool for the association analysis, there are some limitations which should be taken into account. The history of recombination between pairs of SNPs can be estimated with the use of the LD measures (e.g. D’ or r2 ) (Devlin and Rish 1995). If you are using the measure D', the haplotypes are then defined as regions where the standard measure of disequilibrium D' is assumed to be 1 (or still very close to the unit value, making appropriate reference to the confidence intervals), in the absence of recombination, for all (or almost) the pairs of markers in the region. A haplotype is a particular combination of alleles at different polymorphic sites on the same chromosome observed in a population at strong LD. Haplotypes are inherited regions without substantial recombination and the level of LD is very high. In a preliminary analysis, it can be seen that the genetic variation is organized into relatively short segments in strong LD, identified as haplotype blocks, each containing a few common haplotypes separated by hot spots of recombination. You can define hot spots those segments of average length less than 10 kb, where the measure of r2 (the squared correlation between alleles present at two loci) between two closed markers never exceeds the value of 0.10. Hot spots, in essence, represent interruptions of LD, they make up less than 10% of the sequences and occur in about 50% of recombination events. These differences of the recombination rate explain, at least in part, the heterogeneity of LD values between SNPs markers along the genome of several species. The most attractive use of haplotypes is the idea that the common haplotypes can capture most of the genetic variation. A reduced number of SNPs (tag SNP), correlated with the other markers of the same haplotype, can be than used (Stram 2004).

In fish breeding, most production traits are affected by a large number of genes with possible interactions among them. As a consequence, in studies of complex traits, the single-locus analysis does not produce reliable results (Goddard 2009). Another disadvantage of the single-SNP approach is that LD could extend to a wide genomic region. In this case, the detection of the region containing the true mutation and the significant associated SNPs could be difficult (Pryce et al. 2010). A possible solution to this problem could be to fit all SNPs simultaneously by using Bayesian models (Gonzàlez-Recio and Forni 2011). Whatever the method used for GWAS, SNPs declared associated with a trait have to be validated, even if a stringent threshold is used to detect the statistical associations. The best way to validate the detected SNPs is to verify the associations in an independent population. GWAS results depend on several factors such as the actual genetic population structure, genetic characteristics of loci and LD within the population (Meuwissen 2007). LD between markers and QTL depends on the factors such as marker density, the effective size of population, and relationships between the selection candidates and the reference population (Mangin et al. 2012). Increasing the density of SNPs will improve the power of QTL detection (Correa et al. 2016). The results of numerous research studies have shown that the confidence intervals for mapped QTLs are generally large. This situation implies the presence, within the confidence intervals, of a large number of genes (1000) and, therefore, it is difficult to identify the single SNP. Second, the use of markers associated with QTLs is complicated, as this association can vary from family to family. An alternative to this method for the fine mapping of the QTL consists of using linkage analysis together with the LD in the population (Hayes et al. 2009a; Meuwissen et al. 2011).

The GWAS method is best used when re-sequencing and reference sequences and an HD genotyping array are available. GWAS is used to identify single-nucleotide polymorphisms (SNPs) and genomic regions associated with the most important production, physiological and fitness traits. LD pruning is used to select the SNPs. Firstly, the phenotypic data should be normalized. The sample size and the number of SNPs must be adjusted to increase the level of significance. The GWAS method was initially used for the study of human diseases and allowed the identification of several SNP associated with these diseases (Visscher et al. 2017). At the beginning, the main problem with this method was the high rate of false positives. This situation derives from the very high number of statistical tests carried out and the significance threshold (Brzyski et al. 2017). Correction methods used for multiple tests include: reducion of the false discovery rate (FDR) or the Bonferroni method. Another factor that increases the number of false positives is the kinship between individuals in the population (Zhu and Yu 2009). The initial GWAS method was later replaced by the linear mixed model (LMM) which considers the kinship structure of the population (Yu et al. 2006). The population structure can be analyzed using the Structure (Pritchard et al. 2000) or DAPC (Jombart et al. 2010) softwares. Several algorithms are available for calculating the kinship matrix (Van Raden 2008) and false positives are analyzed using a fixed effect (population structure) and random effect (polygenic background) (Yu et al. 2006). The following softwares are available: Emma (Kang et al. 2008), FaSTLMM (mixed linear transformed and spectrally factored model) (Lippert et al. 2011) and Gemma (Zhou and Stephens 2012). In these methods, the computation efficiency is increased as the same variance components are not required, for all the candidate markers (Zhou and Stephens 2012). FaST-LMM uses a subset of SNPs (Lippert et al. 2011) while the Gemma software uses all markers. These methods are now available (multi-locus MLM with random SNP effect) also in the Tassel (Bradbury et al. 2007) and Gapit softwares (Tang et al. 2016).

Most of the research on fish species was carried out using individual SNPs. The alternative is to use haplotypes with 8–15 alleles. GWAS allows to detect associations between genetic markers and production, disease resistence to specific pathogens or functional phenotypes (Goddard and Hayes 2009). These strategies associated with the availability of a high-density reference sequence and a physical map of the genome allow detailed analysis of polymorphisms (Rexroad and Vallejo 2009; Vallejo et al. 2019). RNA-seq and GBS methods are also used for QTL mapping. The availability of high-density SNP arrays has enabled several GWAS studies to be carried out in the Atlantic salmon (Salmo salar) (Tsai et al. 2015) and rainbow trout (Robledo et al. 2017). GWAS was used to study and dissect complex importatant traits (Correa et al. 2015, 2016; Gonzalez-Pena et al. 2016; Tsai et al. 2016b; Gutierrez et al. 2018). QTLs identified for rainbow trout include stress response (Vallejo et al. 2009), spawning time (Colihueque et al. 2010), growth rate (Wringe et al. 2010), osmoregulation capacity (Le Bras et al. 2011), salinity tolerance (Norman et al. 2012) and for Atlantic salmon growth rate (Houston et al. 2009; Baranski et al. 2010; Gutierrez et al. 2012; Tsai et al. 2015; Yoshida et al. 2018), salinity tolerance (Norman et al. 2012), resistence to sea lice (Tsai et al. 2016a), pancreatic disease (Gonen et al. 2015a) and amoebic gill disease (Robledo et al. 2018b). In Table 2 are reported some QTLs identified for the Atlantic salmon and rainbow trout. However, these QTLs explain only a part of the total additive genetic variation of the character under consideration and have only been confirmed in some families/populations. The annual period of sexual maturity is an important trait for salmonids (Küttner et al. 2011). QTLs for sexual maturity have been identified in rainbow trout (Easton et al. 2011), Atlantic salmon (Gutierrez et al. 2014) and Artic charr (Küttner et al. 2011). One of the main QTLs for sexual maturation in Atlantic salmon was found independently in two studies (Ayllon et al. 2015; Barson et al. 2015). Incorporation of the QTL for the infectious pancreatic necrosis (IPN) in the Atlantic salmon resulted in a 47% reduction of the disease in 2009 (Houston et al. 2012; Moen et al. 2015). So far, effective breeding program based on the MAS approach were based on a simple trait in which the main QTL was known (Yue 2014). Most of the traits being selected in the aquaculture sector follow a polygenic model and therefore the study of QTLs in the era of genomic selection is likely of limited value.

Table 2 Examples of QTLs in Atlantic salmon and rainbow trout species

Genomic selection

Recent innovations in the field of molecular biology have allowed greater efficiency and accuracy in animal selection (de los Campos et al. 2013). Meuwissen (2007) defined genomic selection the simultaneous selection of several tens or hundreds of thousands of markers, with which it is possible to cover the entire genome so densely that all genes are expected to be in LD with at least some of the markers used. In particular, genomic selection is based on the prediction of the genomic breeding values ​​of individuals (Genomic Estimated Breeding Values, GEBV), consisting of a large number of DNA markers that are in LD with QTLs encoding traits with important economic value and that span the genome (Solberg et al. 2008; Clark et al. 2011; Börner and Reinsch 2012). The first proven methodology for implementing genomic selection was presented by Meuwissen et al. (2001). Meuwissen et al. (2016), they first introduced genomic selection as a selective strategy based on GEBV obtained by the use of molecular markers. This method implicitly recognizes the fact that quantitative traits are controlled by the segregation of a large number of loci (QTLs) and assumes that the markers are in LD with the QTLs (Calus et al. 2008; Calus and Veerkamp 2011). SNPs are located in coding and noncoding (more frequently) regions, and we assume that each quantitative trait is controlled by approximately 300-700 QTLs (Hayes et al. 2009b). Therefore, the higher the density of the markers and the stronger the level of LD between markers and the QTL, and, consequently, the proportion of genetic variation that can be explained through the markers. Genome-wide SNPs are used to calculate the genomic relationship matrix (GRM) and the estimated genomic values (GEBV) (Van Raden et al. 2009; Aguilar et al. 2011; Goddard et al. 2011; Legarra et al. 2014).

Genomic selection allows for greater accuracy in EBV estimation than traditional methods (Habier et al. 2007; Forni et al. 2011; Vallejo et al. 2017). However, the accuracy of the GBLUP estimate varies mainly in relation to the heritability of the selected trait (Daetwyler et al. 2008), allelic substitution effect for the QTL (Meuwissen et al. 2001; Goddard 2009), and the size of the training population (Daetwyler et al. 2008; Andonov et al. 2017).

Genomic evaluation is a two-step method: (i) 1st step: the genome is divided into chromosome segments. A reference population (1000+ individuals, genotypes and phenotypes) is used to estimate the SNP or haplotype effects at each chromosome region; (ii) 2nd step: selection candidates are genotyped, and their GEBVs are calculated as the sum of segment effects. Phenotypic and pedigree data are used for the estimation of the SNP effects in the reference population (Fig. 1). Evaluations of marker effects in the reference population should be updated at each generation. The population is generally divided into reference and validation groups for the validation of statistical models (Eggen 2012; Wu et al. 2015). The population used for training must be adequate, especially for traits with low heritability. Subsequently, the effects of SNPs are validated in a population independent of the training population and verified for accuracy. All the segment effects are fitted simultaneously, reducing the bias.

Fig. 1
figure 1

Calculation of GEBV in a genomic selection scheme

If all the marker effects show a normal distribution and identical variance, the accuracy of GEBVs is equal to the traditional Best Linear Unbiased Prediction (BLUP) animal method. This method considers all markers to estimate the genetic value of the candidate for selection. These effects can also be used to select animals whose phenotype is unknown (Meuwissen et al. 2001). The BLUP method uses information from the animals' pedigrees and phenotypes to estimate breeding values (EBVs). The use of molecular markers (SNPs) provides a new source of information, which improves the accuracy of EBV prediction. We expect the selection response will be highest for traits with low heritability and in those that cannot be measured in the candidate for selection (such as disease resistance) (Taylor 2014). For markers with larger effects (e.g. longer tails) and different variances, SNP-BLUP and G-BLUP methods were first used (de Roos et al. 2011; Habier et al. 2011). The SNP-BLUP (RR-BLUP) model assumes that each SNP has a very small effect on the genetic variance of the trait. The GBLUP method computes EBVs using traditional mixed model equations, and the relationship matrix is calculated as relationship coefficients among the observed genotypes (Van Raden 2008; Hayes et al. 2009a). The G-BLUP and SNP-BLUP methods assume that all SNP effects are non-zero, small and normally distributed (Dong et al. 2016). After the GBLUP method, the ssGBLUP (single step) method was developed (Aguilar et al. 2010; Legarra et al. 2014; Yoshida et al. 2019). In this method, all the information (phenotypic, genotypic and genealogical) is simultaneously analysed (Aguilar et al. 2011) weighting the genomic and pedigree relationships. The two methods (GBLUP and ssGBLUP) assume that all SNPs are equal (weight and variance).

Subsequently, alternative methods were developed, such as BayesA, BayesB, BayesC and Lasso, using different weights of SNPs (high LD with a mutation or associated with a QTL) and the presence of QTLs) (Meuwissen et al. 2001; Habier et al. 2011; Gianola 2013; Fernando et al. 2014), WGBLUP and WssGBLUP (Snelling et al. 2011; Tiezzi and Maltecca 2015) and TABLUP (Zhang et al. 2010). Bayes A is set up at two levels. The first is similar to the SNP-BLUP model. The second is organized at the level of variance across the SNPs. In Bayes B, some SNPs are located in regions with no QTL (zero value), whereas some show a moderate or large effect because they are in linkage disequilibrium with QTLs. Lasso Bayes uses a similar model of Bayes A to estimate marker effects, but it makes a different assumption concerning the distribution of marker variance (sampling from a double exponential distribution). Bayesian methods allow us to obtain a comparable accuracy compared to traditional methods by means of a shrinkage of the segment effects based on prior QTL distribution knowledge (Habier et al. 2011).

Using SNP low-density bead chips, few differences in accuracy between GBLUP and Bayesian methods were observed (Habier et al. 2011). The genomic evaluation explains a part of the quantitative variance (between 40 and 60% depending on the traits) (Schenkel 2009). Total genetic effects are then estimated by combining genomic evaluation and polygenic effects (Fig. 2). A single or multi-step method can be used to combine genomic and polygenic information (Calus and Veerkamp 2011).

Fig. 2
figure 2

Conceptual view of genomic evaluation

Traditionally, the selection of the best animals is made using data collected over time on its own performance, records of relatives and relationships across populations. This approach means that for traits measured only in older age, the generation interval will increase. The accuracy of evaluations is almost equivalent for males and females (Schaeffer 2006). Several studies have shown some benefits of including males and females in the reference population (Fig. 3) (Sonesson and Meuwissen 2009; Toro et al. 2017). The use of genomic data keeps the response to selection constant over generations (Sonesson 2007; Sonesson and Meuwissen 2009). Solberg et al. (2008) reported that to obtain the same accuracy in genomic selection, it is necessary to use a greater number of SNPs (at least three times) than microsatellite markers and to increase the density of markers from 1Ne(effective population size)/morgan to 8Ne/morgan. For a family of 100 individuals, at least 24,000 loci are required. The response to genomic selection using a cohort of 100 families improves significantly compared to the use of single families (Khatkar et al. 2017). In this case, the cost of infrastructures used for breeding decreases significantly compared to traditional systems. Multistage selection (phenotypic + genotyping) is probably the optimal solution (Khatkar et al. 2017).

Fig. 3
figure 3

Accuracy of EBVs in the reference population for different heritabilities (adapted from D’Agaro et al. 2007)

Optimisation of fish breeding schemes using genomic selection methods

Genetic progress in fish breeding depends on several factors, such as selection intensity, generation interval and accuracy of genetic evaluations (Täubert et al. 2011; Mrode 2014). Advances in molecular genetics, which became available in recent decades, allowed us to expand, or even to rethink, the scenario of fish selection. These new discoveries have allowed us to identify the chromosome position (within a confidence interval) of quantitative trait loci (QTLs) responsible for relevant variation in several traits of economic interest and functional traits (Dekkers and Hospital 2002). For example, Ozaki et al. (2007) used microsatellite markers to identify two QTLs that influence rainbow trout (27 and 34% of the phenotypic variation) resistance to infectious pancreatic necrosis (IPN), and Rodriguez et al. (2004) developed three different linkage maps. The most famous example in fish breeding is the QTL which encodes resistance to IPN in Atlantic salmon (Houston et al. 2008; Moen et al. 2009; Moen et al. 2015). The higher frequency of the allele at this QTL induces almost total resistance to IPN, and a breeding plan was rapidly implemented in Norway. However, it appeared from these studies that only some traits are determined by important QTLs (Haidle et al. 2008; Gutierrez et al. 2012). Another successful example is the selection of populations of rainbow trout resistant to bacterial cold water disease (Bcwd) using a 57K SNP chip (Vallejo et al. 2017; Liu et al. 2018). Unfortunately, the direct use of genotypic information in MAS programmes (Dekkers 2007; Tiezzi and Maltecca 2015) or practical applications were limited. These results can be partly explained in terms of the large number of loci controlling the quantitative traits and the limited number of markers available. The reduction in the accuracy of EBVs is partly explained by the fact that QTLs are not identified directly by specific SNPs, but by the associated markers. In more recent years, new methods have been discovered for the simultaneous determination and precise genotyping of hundreds of thousands of SNP markers using new SNP HD arrays (Houston et al. 2014; Yáñez et al. 2014). SNPs, although less informative than other markers (for example, microsatellite markers), allow efficient monitoring, and, at lower cost, genetic variations among many samples (Van Tassell et al. 2008). Recombination occurring between genes can reduce the LD level at each generation. If LD (between the marker and a QTL) is present only within families (and not between families), the recombination occurring between genes can break the association (Wientjes et al. 2013). This situation makes the MAS unattractive because the LD must be verified at each generation for each family (Dekkers 2007). MAS schemes can be used only when the QTL explains a large proportion of total trait variance (as is the case for IPN resistance in Atlantic salmo (Houston et al. 2008, Laghari et al. 2014). The discovery and annotation of thousands of new genomic variants allow the production of new SNP arrays, and, in this way, the implementation of effective genomic programmes (Abdelrahman et al. 2017; Guppy et al. 2018). Currently, there are several SNP arrays available for salmonid species. Houston et al. (2014) identified approximately 132 K SNP markers in different populations of farmed and wild Atlantic salmon using a SNP array produced by Affymetrix (Santa Clara, CA, USA), and Yáñez et al. (2014) validated 160 K SNP markers using a 200 K array for Atlantic salmon (for wild and farmed populations in Europe, North America and Chile). A 57 K SNP chip (Affymetrix Axiom, Santa Clara, CA) is available for rainbow trout (Palti et al. 2014).

The discovery in the last decades of the genomic selection approach has completely changed the method of calculation of the breeding values (GEBV) for the most important domestic livestock species (Meuwissen et al. 2016), and, potentially can do the same in the aquaculture sector (Correa et al. 2017). Recent studies have shown that the genomic selection method can be effectively used to improve some salmonid production and disease resistance traits (Yáñez et al. 2016; Yoshida et al. 2018), and the first example of a genomic selection programme was established for the genetic improvement of Atlantic salmon (Houston et al. 2014; Ødegård et al. 2014; Tsai et al. 2015). In recent years, several aquaculture breeding programmes have applied the genomic selection method: Atlantic salmon, rainbow trout, Nile tilapia and some crustacean species (Pacific whiteleg shrimp (Litopenaeus vannamei)) (Norris 2017; Zenger et al. 2017, 2019). While routine implementation of genomic selection is now largely carried out in salmonids (Yáñez et al. 2014), it can be expected that in the near future, this method will progressively spread to other fish species. However, genomic selection is an expensive method (fish genotyping costs range between 40 and 80 €/sample), so similar to the MAS method, it will be mostly relevant for traits of high economic value. Genomic selection has so far been applied for bacterial cold-water disease resistance in rainbow trout (Abdelrahman et al. 2017, Vallejo et al. 2017, Liu et al. 2018; Yoshida et al. 2018) and sea lice resistance in Atlantic salmon (Tsai et al. 2016b, 2017; Robledo et al. 2018b). In different studies, the accuracy of the GEBVs varied from 0.16 to 0.83 for different traits (+ 24% for growth traits and 22% for disease resistance) compared to traditional methods based on genealogy (Ødegård et al. 2014; Tsai et al. 2015; Robledo et al. 2018a, 2018b). Bayesian models have shown results comparable to traditional GBLUP methods (Zenger et al. 2019). Recently, Barría et al. (2018) reported an increase in accuracy compared to traditional methods of 155%.

Genomic selection can increase the selection response in the fish breeding sector (+10% for body weight) compared to traditional selection methods (Campos-Montes et al. 2013; Ødegård et al. 2014; Castillo-Juárez et al. 2015). In the future, the use of young breeders (reducing the generation interval), the reduction of the inbreeding rate (Vandeputte and Haffray 2014; Dupont-Nivet et al. 2006; Holtsmark et al. 2008) and the incorporation of dominance and epistatic effects (Dupont-Nivet et al. 2008; Nilsson et al. 2016) may further increase the genetic response (Campos-Montes et al. 2013; Castillo-Juárez et al. 2015). Improved genome annotation techniques and RNA sequencing (RNA-seq) allow the identification and characterization of new SNPs. The high fecundity and the family sizes used in fish breeding plans allow us to obtain a high selection accuracy (because of the close relationship between the reference population and selected candidates). The optimal size of the population (genotyped and phenotyped > 1000 individuals) and the population structure are important factors determining the success of the genomic selection scheme (Daetwyler et al. 2010a, 2013). A density of SNPs equal to 1000–2000 is sufficient to obtain excellent results (Zenger et al. 2019; Kriaridou et al. 2020). GEBV accuracy decreases dramatically in each generation if the relationship between the reference population and selected animals is low (Tsai et al. 2016b). Genomic selection is applied to raise fish in a common environment and thereby to reduce costly infrastructure interventions (Fernández et al. 2014).

Genomic selection using full-sib families can be performed using low-density SNP arrays to estimate GEBVs within known pedigrees (Lillehammer et al. 2013; Tsai et al. 2016b; Tsairidou et al. 2020). The use of low-density SNP arrays (GT-Seq) (Campbell et al. 2015) or fluorescence-based arrays reduces the costs of genetic analyses, particularly in aquaculture, where thousands of potential selection candidates are used (Lillehammer et al. 2013; Gutierrez et al. 2018). These technologies can be easily integrated using the imputation technique (Tsai et al. 2017; Yoshida et al. 2018; Tsairidou et al. 2020). In this case, only a sample of animals is genotyped with a SNP HD array with a significant reduction in costs (Tsai et al. 2017). Several fish breeding programmes routinely use low-density SNP arrays for parental assignment (Vandeputte and Haffray 2014); therefore, in this case, genomic selection can be applied with a small additional cost. It is worth noting that, however, the use of SNP HD arrays, complete genome sequences and amplicon genotyping allow an increase in the accuracy of the selection process. Whole-genome sequencing of selected individuals (e.g. parents) and the imputation method can improve the accuracy of the estimate (Badke et al. 2013; Browning and Browning 2016; Bilton et al. 2018). The accuracy of the imputation depends on several factors: the number of markers used, LD at the local level, minor allele frequency, and correlations between individuals (reference and selected) (Druet et al. 2010; Hickey et al. 2012). The accuracy of imputation has been measured in a few fish breeding programmes. Kijas et al. (2017) found an accuracy in Atlantic salmon of 0.89-0.97, and Tsai et al. (2017) equal to 0.90. Note that, to obtain an accurate imputation process, the precise location of genetic markers in the genome is needed. For many fish species, this information is still incomplete (Abdelrahman et al. 2017).

In fish breeding programmes, the cost of collecting phenotypic data may, in some situations, present a limitation to achievable genetic progress (Føre et al. 2018). The use of computer vision technologies such as video cameras or sensors could be helpful (Zenger et al. 2019; Saberioon et al. 2017) to monitor and to record certain production, behavioural and health characteristics of fish. Some examples are reported in the literature concerning fillet analysis by means of the hyperspectral imaging method (Saberioon et al. 2017) or camera monitoring of sea lice in Atlantic salmon farms (Føre et al. 2018). Recently, automatic food distribution systems and the use of video cameras have allowed for faster collection of phenotypic data in fish farms (McCarthy et al. 2010; Cobb et al. 2013; Saberioon et al. 2017). Machine vision systems (MVS, 2D or 3D imaging) allow the acquisition of various phenotypic measurements, such as the shape, size, weight, colour and quality of fish fillets (Schmidhuber 2015; Saberioon et al. 2017). The accuracy of the estimates is very high (≥ 0.95). Miranda and Romero (2017) measured body lengths in rainbow trout using MVS systems, and Colihueque et al. (2011, 2017) compared the fillet colour between a traditional system and a spectral model produced by MVS in Atlantic salmon and rainbow trout. Additional techniques can be used in data collection, such as hyperspectral imaging (HSI) (spectroscopy and imaging) and near-infrared (NIR) spectroscopy (Liu et al. 2013). These techniques allow us to determine the chemical characteristics (proximate analysis) and physical characteristics (qualitative characteristics of the carcass) with high precision (r > 0.8) (Saberioon et al. 2017). These technologies are automatable and non-invasive. The use of image analysis software based on neural network (ANN) algorithms can be used for the rapid and automated phenotyping of animals (Grys et al. 2017).

Prospects and challenges

Since the start of the “genomic era”, many innovations and advantages in the field of fish breeding have become available (D'Agaro 2018). On the other hand, there are some downsides. The first difficulty seems to be associated with the large amount of new data that genomic selection requires. In fact, the possibility of combining different sources of information makes the interpretation work very complex. To date, new genomic methods have enabled the integration of the information provided by molecular genetics with those hitherto available in the field by quantitative genetics. Based on these results, several breeding programs have already started to use genomic selection methods based on GEBVs.

Population genomic data sets have also provided valuable information to address many aspects of the eco-evolutionary history in several salmonid species, including the characterisation of population structure, local adaptation and genetic diversity (Crespi and Fulton 2004; Hohenlohe et al. 2011; Lamaze et al. 2012; Bourret et al. 2013; López et al. 2014; Yáñez et al. 2015; Liu et al. 2017; Lopez Dinamarca et al. 2019). Genomic information that is shared among individuals in a population can be used to construct a genomic relatedness matrix (Gienapp et al. 2017). The availability of thousands of SNPs also has allowed the development of conservation strategies in several countries (Abadía-Cardoso et al. 2013; McMahon et al. 2014; Bradbury et al. 2015; Bohling et al. 2016). Hand et al. (2015) used 10,267 SNPs to study the introgression of alleles of rainbow trout in wild populations of Cutthroat trout (Oncorhynchus clarkii lewisi) and to identify chromosomal regions affected by selection. Currently, such approaches have only rarely been implemented in brown trout (Sušnik Bajec et al. 2015). The identification of epigenetic traces in fish allows a better understanding of genotype-phenotype interactions (Gavery and Roberts 2017). New technologies allow us to understand some mechanisms, such as cytosine methylation and histone modifications. These tools allow us to understand how some active SNPs work in specific environmental conditions (Gutierrez et al. 2016) and to identify the signatures of selection (Kjærner-Semb et al. 2016). Comparative genomics and orthological analysis of wild populations can be used to identify and to analyse SNPs of interest (Emms and Kelly 2015).

In the future, additional novel technologies such as gene editing can be used to modify genes by adopting new targeted methods. These methods are based on engineered nucleases (Crispr-Cas9; Crispr/ScCas9; and Crispr-Cas12) (Zhang et al. 2018a, 2018b; Gosavi et al. 2020), which can generate a break in the double strand of DNA in a precise sequence. When cellular mechanisms repair the break, they can cause insertions or deletions that result in inactivation of the gene or replacement with an alternate sequence if an appropriate template of donor DNA is provided. The main interest in these technologies is because they allow a high success rate on targeted DNA modifications, even in homozygosity. This method can greatly reduce the time required to obtain a line with the desired characteristics, overcoming the ethical implications, as in the case of introducing a “natural” mutation from another population at the same species (Gratacap et al. 2019). At present, this editing methodology in aquaculture has been used to obtain albino and sterile Atlantic salmon (Wargelius et al. 2016), common carp with a reduced number of spines and more muscle cells (Zhong et al. 2016) and sex-reversed Nile tilapia (Li and Wang 2017). While none of these experiments have reached the commercial production stage, these modern biotechnological methods clearly provide a new impetus to the direct targeted gene modification approach. The commercial applicability in the genetic improvement of fish species will, however, depend on how these animals are considered by the regulatory authority (Abdelrahman et al. 2017).


The genetic improvement of fish has historically been carried on the base of estimated breeding values of animals, starting from their performance and pedigrees. Currently, genomic selection has radically changed this scenario, allowing us to estimate the genomic value of an animal even without knowing its pedigree and at birth. This information then allows us to speed up the selection process and increase the accuracy of estimates. In salmonid species, several projects have concentrated their activities on the collection of a large number of genetic markers and have increased the reference population. In recent years, the accuracy of the GEBV of selection candidates has rapidly increased. Future strategies will allow simultaneous management of the use of genomic information and control of inbreeding in these and in new species.