Genetic analysis of swimming performance in rainbow trout (Oncorhynchus mykiss) using image traits derived from deep learning

Abstract

The physical and physiological condition of fish directly influences their swimming performance, which is crucial for their health and survival. This study explored how physical characteristics affect swimming performance in rainbow trout. 3D images were used to capture the morphology of fish and assess its impact on critical swimming speed (U_crit), measured via individual swim tests. A convolutional neural network (CNN) was utilized to predict U_crit from the images. Using Gradient-weighted Class Activation Maps (GradCAM), image regions that contributed to U_crit predictions were visualized. These regions were further refined into areas that are biologically relevant to U_crit, leading to the definition of four swim traits: head volume, caudal fin volume, epaxial muscle volume, and shape. Our findings indicated that U_crit is moderately heritable. Genetically, heavier fish demonstrated poorer swimming performance; among fish of the same weight, those with larger and broader epaxial muscles, larger heads, and smaller caudal fins performed worse. Although genetic improvement of U_crit is feasible, caution is advised because of potential correlated responses that reduce the body volume and epaxial muscle volume. The interdisciplinary workflow (data collection, model construction, visualization, interpretation, definition, and evaluation) in this study demonstrated how image-based deep learning can be used as a hypothesis-free approach to deepen the understanding of the genetic background of complex traits. Additionally, it highlights the value of genetic analysis to validate the physiological interpretation of Explainable AI, broadening the opportunities to discover novel phenotypes in aquaculture.