High-throughput NIR-based phenotyping for genomic prediction of feed intake and production traits in Atlantic Salmon

Sammendrag

Accurate and large-scale phenotyping of feed intake and growth traits is crucial for genetic improvement and ensuring economic sustainability in fish breeding programs. Although standard methods for measuring growth traits such as body weight (BW) and average daily gain (ADG) are relatively high-throughput, conventional techniques for assessing individual feed intake (IFI) and feed efficiency (FE) remain inefficient, labor-intensive, time-consuming and costly. This limits the ability to conduct routine large-scale phenotyping. Here, we validate non-contact interactance near-infrared spectroscopy (iNIR) as a rapid, cost-effective and non-invasive phenotyping tool for predicting key production traits, including IFI, BW, ADG, residual feed intake (RFI), and feed conversion ratio (FCR), in Atlantic salmon. Using a cohort of 680 salmon smolts, we compared different supervised learning techniques, including Lasso, partial least squares regression (PLSR), Bayesian learning and convolutional neural network (CNN) for trait prediction. Lasso regression achieved the highest accuracy for IFI (r = 0.78 ± 0.02), BW (r = 0.85 ± 0.02), and ADG (r = 0.85 ± 0.02), while PLSR performed best for RFI (r = 0.28 ± 0.04) and FCR (r = 0.38 ± 0.03). We then conducted genomic validation of the iNIR predictions to ensure their reliability for use in breeding programs. First, we estimated the genomic parameters for both the reference measurements and iNIR-predicted traits using the restricted maximum likelihood (REML) approach. Subsequently, we carried out genomic prediction through 20 random iterations of a 5-fold cross-validation scheme to validate the utility of iNIR-predicted phenotypes in the genomic prediction model. Our analyses revealed moderate to high heritability for iNIR-predicted traits, ranging from 0.31 to 0.50. Additionally, iNIR-predicted traits exhibited strong genetic correlations with reference measurements, ranging from 0.67 to 0.96, and demonstrated robust genomic prediction stability, ranging from 0.83 to 0.91. These encouraging estimates confirm that iNIR-predicted traits can serve as reliable proxies. Furthermore, incorporating iNIR-predicted traits into multi-trait genomic prediction models enhanced predictability by 14 to 20 % and stability by 4.9 to 8.5 %, compared to single-trait models for IFI, BW and ADG. While predictions for RFI and FCR were less accurate, likely due to their metabolic complexity, our results demonstrate that iNIR can be a viable high-throughput and cost-effective phenotyping tool for growth and feed intake, thereby advancing genomic selection in Atlantic Salmon. Although the phenomic predictions showed room for improvement, genomic validation highlighted the potential of using inexpensive and non-destructive iNIR proxy traits to improve genomic selection. In practice, genomic reference cohorts comprising a few thousand individuals can be genotyped and phenotyped for reference traits that are both expensive and difficult-to-measure phenotypes (such as IFI) alongside rapid, inexpensive iNIR to generate accurate iNIR proxies. Subsequently, larger cohorts, including selection candidates or informant fish on the slaughter line, can be phenotyped with iNIR to improve multi-trait genomic predictions in selection candidates. However, independent studies replicating these findings are needed.