Item

Abstract

Animal embryos stereotypically produce body plans with species-specific features and appearance, but the developmental dynamics to achieve appropriate growth and form differ widely between species. Accurate assessment of morphogenesis and developmental tempo is therefore central for quantitative studies of embryonic development, but given the fluid nature of embryogenesis as well as considerable variation between individuals, manual assessment remains subjective and challenging. To overcome this challenge, we have developed TwinNet, a neural network that allows to quantitatively analyze developmental dynamics in diverse model organisms using artificial intelligence and similarity calculations. We first created a dataset of thousands of images of zebrafish development and applied TwinNet to automatically assess embryonic age without the use of pre-defined stages. In addition, TwinNet allowed us to calculate developmental tempo and variations between individual embryos. We then extended our method to medaka, three-spined stickleback, and C. elegans, demonstrating that TwinNet can be robustly applied in different species even when only limited data is available. Furthermore, TwinNet was able to automatically generate atlases of the main developmental epochs in diverse species, which quantitatively showed that development is characterized by the alternation of periods, in which embryonic morphologies change, and phases, in which embryonic morphologies undergo little change. In summary, TwinNet can be used as standardized approach with various possibilities for objective, quantitative and multiparametric studies of embryonic development. The code of our software as well as terabytes of training and evaluation data are freely available from https://github.com/mueller-lab/TwinNet and https://doi.org/10.48606/50, and the method is open to further development by the community.