Deep learning approach for genetic selection of stress response in the Amazon fish Colossoma macropomum

Abstract

This study examined skin color variation in tambaqui (Colossoma macropomum) in response to stress, focusing on morphological and physiological color change mechanisms. A computer vision system (CVS) based on the DeepLab V3 model with ResNet-50 was developed to automate countershading intensity detection. Images from 3780 fish across two populations were used to train a model and estimate genetic parameters for countershading intensity. Morphological color changes were induced in confinement tanks, with countershading intensity observed after 10 days. Physiologically, the α-MSH hormone expanded melanophores by 80 %, intensifying countershading. The CVS achieved high accuracy (88.2 %) for large-scale phenotyping, with moderate to high heritability estimates for color phenotypes: 0.456 ± 0.122 for black pixel percentage, 0.494 ± 0.128 for mean pixel intensity, and 0.192 ± 0.059 for the number of pixels. Low correlations with growth traits suggest that countershading selection can occur without affecting growth, highlighting its potential in breeding programs to improve appearance and stress resilience.