Atlantic salmon is the prime aquaculture species in Norway with high global demand and export value. Most of the salmon production cost goes to feed and thus improving feed efficiency (FE) is crucial to reduce this cost and environmental impact. However genetic studies on feed efficiency are limited, primarily due to difficulty in phenotyping individual feed intake. This study aims to overcome this limitation by (i) quickly and accurately phenotyping individual feed intake under commercial conditions; (ii) training deep learning model for efficient object (bead) detection and (iii) analyzing the genetic architecture of FE traits including individual feed intake (FI), average daily weight gain (ADG), Residual feed intake (RFI), and feed conversion ratio (FCR).
Atlantic salmon belonging to the 2021-year class of MOWI genetics, Norway underwent digital phenotyping experiments. The snapshots of individual FI were recorded using X-ray imaging and a deep learning model (YoloV5) for bead detection was trained to estimate the FI accurately from images. The model performed exceptionally well with an R2 of 0.99, low RMSE and a slope of ~ 1 across training, validation and test sets.
FE-related traits showed varied potential for selective breeding with intermediate to high heritability estimates for FI (h2 = 0.20 ±0.05 - 0.50 ±0.06) and ADG (h2 = 0.47 ±0.07 - 0.54 ±0.06), and low to moderate for FCR (h2 = 0.08 ±0.04 - 0.23 ±0.06) and RFI (h2 = 0.10 ±0.05 - 0.17 ±0.11). We are currently diving deep into GWAS and genomic prediction to better understand the genomics of FE, and more results will be ready for the conference.