A high-throughput phenome-based analysis of morphological variation and environmental adaptation in extremely high-altitude schizothoracine fishes

Schizothoracine fishes are distributed in the Nagqu region, which is the hinterland of the Tibetan Plateau. They have adapted to the cold and strongly ultraviolet environment of the plateau and show diverse morphologies, which makes them ideal for studying the relationship between phenotype and environment. However, traditional morphological measurements are time consuming and labor costly. In this study, we propose a deep-learning-based method for acquiring high-throughput phenotypic data of fishes, including image dataset construction of schizothoracine fishes (including front, side, and top views), fish keypoint detection based on the You Only Look Once (YOLO) model, and reconstruction of 2D and 3D keypoint coordinates based on multiple views. A total of 7050 phenotypic data points consisting of keypoint distances and angles were constructed for each fish and were highly correlated (>0.98) with the corresponding data measured manually. We obtained phenotypic data for a total of 525 fishes from three schizothoracine fish groups inhabiting riverine, lacustrine, and river–lake transitional environments using the proposed phenotypic data acquisition method. We employed a random forest algorithm to classify the groups, achieving a classification accuracy of 96%, and identified 15 morphometric indices that exhibited statistically significant differences., of which 6 were related to head morphology, 6 related to body shape, and 3 related to tail morphology, based on the random forest algorithm. Specifically, river-living schizothoracine fishes showed a blunt head, robust body, and elongated caudal peduncle which may reflect adaptations to the turbulence of the river, while the lake-living schizothoracine fishes have the opposite effect. Schizothoracine fishes at the river–lake transitional zones were phenotypically characterized as being in the middle of the two phenotypes, and these presumably reflect adaptations to their lake habitat. This study provides a methodological reference for obtaining high-throughput phenotypic data on fish and a theoretical basis for understanding the adaptation of very high-altitude schizothoracine fishes to their environment. • We used computer vision technology to efficiently and intelligently acquire high-throughput phenotypic data from three views of the fish body. • We used machine learning to filter important indicators of morphological differences between habitats, with a classification accuracy of 95%.