Introduction
Flavobacterium columnare, the aetiological agent for columnaris disease (CD), is a major challenge for several aquaculture species, including rainbow trout (Oncorhynchus mykiss). The disease manifests via acute and chronic infections and, leading to high mortality rates, particularly in juvenile fish (Declercq et al., 2013; Evenhuis et al., 2014). Due to the impact of the disease, several vaccines have been developed to enhance CD immune response (Malecki et al., 2021); however, to date no vaccine is commercially available for salmonids. As a result, current practices rely on antibiotic and antimicrobial treatments, posing a risk to both the environment as well as the development of microbial resistance. In response to these challenges, selective breeding for disease resistance offers a viable alternative to improve the sustainability and welfare of farmed populations (Houston et al., 2020). Previous studies on trout CD resistance revealed substantial genetic component (0.20-0.43) and moderate genetic correlation (~0.62) between distantly related populations (Calboli et al., 2022; Fraslin et al., 2022). In addition, performance of genomic prediction on a CD challenge population revealed 18.6% improvement in prediction accuracy compared to application of pedigree-based approaches (Fraslin et al., 2022), highlighting the potential of implementing genomic approaches in selective breeding. While genomic information improved prediction accuracy within a single population, the possibility of assessing disease response using information from unrelated or distantly related populations is currently unknown. The use of previously challenged populations in prediction modelling provides substantial benefits for selective breeding programmes by reducing the need for repeated disease challenge experiments, thereby lowering phenotyping costs and decreasing the number of animals exposed to infectious challenge trials. The aim of the current study was to assess the accuracy of genomic prediction for rainbow trout CD resistance using two distantly related populations.
Materials and methods
Population and genotype filtering
We used publicly available genotype and phenotype information from two commercial rainbow trout populations that participated in a CD survival challenge. Both populations are described in detail in Fraslin et al. (2022) and Calboli et al. (2022). In particular, 1466 fish from the Savon Taimen (ST) population and 2874 fish from the Luke (LK) breeding population with survival records from a Flavobacterium columnare challenge trial were genotyped using a 57K Axiom Trout Genotyping SNP array. Genotypes for all fish were filtered using Plink v2 based on the following criteria: minor allele frequency over 5%, SNP genotype call rate over 98%, genotype missingness less than 2% and Hardy-Weinberg equilibrium deviation (p.value < 10-6). Finally, the physical coordinates of high-quality SNPs were mapped against the latest version of the rainbow trout genome reference (GCA_013265735.3).
Estimation of genetic parameters
Variance components and heritability for each population were estimated based on genomic relationships using a univariate linear mixed model in AIREMLF90, where binary survival was used as the phenotype, the genomic relationship matrix among candidates was fitted as a random effect and the individual sex and tank (3 classes per population) were fitted as fixed effects. Best linear unbiased predictions for each population were estimated using BLUPF90 and the same linear mixed model was applied in a 10-fold cross validation approach. Cross-population prediction of breeding values was performed with the same model and software as before, using the individuals of one population to train the model and the individuals of the second population for validation. The process was iterated to assess prediction accuracy in both populations. Across all prediction scenarios, prediction accuracy was estimated as the correlation between estimated breeding value and true phenotype, divided by the square root of heritability.
Results and discussion
Quality filtering retained a total of 1459 and 2819 animals for ST and LK populations, respectively, with phenotypes for binary disease survival and genotype information for 22,857 SNPs. Disease survival was estimated at 61.5% and 51.2% for the ST and LK populations, respectively. Estimation of heritability within each population using the cross-validation testing approach showed substantial genetic component (0.20 ± 0.05 and 0.19 ± 0.03 for ST and LK, respectively), consistent with previous findings (Calboli et al., 2022). In addition, prediction accuracy using the same approach was moderately high for both the ST (0.66 ± 0.01) and LK (0.71 ± 0.01) populations (Table 1).
Interestingly, accuracy varied markedly between the two cross-population prediction models. In particular, using the LK population to train the prediction model yielded substantially higher accuracy compared to the model trained using the ST population (0.53 and 0.27, respectively). As the latter population included 1459 all-female individuals (offspring of neo-male female ancestors), differences in prediction accuracy between the two populations could reflect sex-specific effects. At the same time, differences in sample size between the two populations (the LK population contains twice as many individuals) could be additionally contributing to the observed discrepancy in cross-population prediction accuracy. Nevertheless, our study underscores the potential in using distantly related populations to predict CD resistance in rainbow trout. Building on these findings, we are currently investigating the use of whole-genome sequence (WGS) data, functional annotation, and complex genetic variation such as structural variants (SVs) in genomic prediction models to evaluate their potential to further improve across-population prediction accuracy. Results of these analyses will additionally be presented at the conference.
Table 1. Accuracy of genomic prediction models assessed using two genetically distant rainbow trout populations. Within-population accuracy was evaluated using a 10-fold cross-validation approach, whereas in the across-population approach, one population was used to train the model and the second was used for validation.
Training population
Method
Prediction accuracy
Savon-Taimen cross-validation
0.66 ± 0.01
Luke cross-validation
0.71 ± 0.01
Savon Taimen across-population
0.53
Luke across-population
0.27
References
Calboli, F. C. F. et al. (2022). Conserved QTL and chromosomal inversion affect resistance to columnaris disease in 2 rainbow trout (Oncorhyncus mykiss) populations. G3 Genes|Genomes|Genetics, 12(8). https://doi.org/10.1093/g3journal/jkac137
Declercq, A. M. et al. (2013). Columnaris disease in fish: a review with emphasis on bacterium-host interactions. Veterinary Research, 44(1), 27. https://doi.org/10.1186/1297-9716-44-27
Evenhuis, J. P. et al. (2014). Early life stage rainbow trout (Oncorhynchus mykiss) mortalities due to Flavobacterium columnare in Idaho, USA. Aquaculture, 418-419, 126–131. https://doi.org/10.1016/j.aquaculture.2013.09.044
Fraslin, C. et al. (2022). Genome-wide association and genomic prediction of resistance to Flavobacterium columnare in a farmed rainbow trout population. Aquaculture, 557, 738332–738332. https://doi.org/10.1016/j.aquaculture.2022.738332
Houston, R. D. et al. (2020). Harnessing genomics to fast-track genetic improvement in aquaculture. Nature Reviews Genetics, 21(7), 389–409. https://doi.org/10.1038/s41576-020-0227-y
Malecki, J. K. et al. (2021). Bioeconomic Analysis of Flavobacterium columnare Vaccine Pond Trials with Channel Catfish. North American Journal of Aquaculture, 83(3), 207–217. https://doi.org/10.1002/naaq.10191