Animal development is a complex process transforming a fertilized egg into a mature adult, with species-specific features emerging during embryogenesis. Traditional methods rely on manual microscopic observation to create staging atlases, which are idealized representations and struggle to capture the nuances of real-world development. Embryos exhibit variations in morphology even within the same nominal stage due to factors like imaging conditions, individual differences, and environmental influences. Furthermore, developmental tempo is influenced by environmental factors such as temperature, making it challenging to establish a standardized approach for comparing developmental stages across different conditions and species. While existing computer-driven methods have been proposed, they often rely on supervised machine learning and require extensive annotated datasets, limiting their generalizability. This study addresses these challenges by proposing a novel, unbiased deep learning approach based on Twin Networks to analyze developmental time and tempo.

The authors review existing methods for staging embryos, highlighting their limitations. Traditional methods rely on manual observation and idealized images, which do not fully capture the natural variation and smooth transitions in embryonic development. While previous studies have explored computer-assisted methods, these frequently depend on supervised machine learning, requiring large, pre-annotated datasets. The limitations of these approaches, particularly regarding their generalizability across various species and experimental conditions, motivate the development of a more robust and unbiased approach.

The core of the methodology is a Twin Network, a deep learning architecture consisting of two identical parallel neural networks that learn hidden representations (embeddings) of input images. The networks calculate similarities between images based on their embeddings, providing a measure of phenotypic similarity. The researchers used a high-throughput imaging pipeline to create a large dataset of over three million zebrafish embryo images. A ResNet101 model was trained for image segmentation and embryo detection. The Twin Network was trained on image triplets (anchor, positive, negative) using triplet loss to learn phenotypic features. The similarity between a test embryo image and a reference set of images across time was used to predict the developmental stage of the test embryo. The method was applied to study temperature dependence in zebrafish and medaka, analyze natural variability in zebrafish development, detect drug-induced phenotypic changes, and automatically generate developmental staging atlases for various species. The Arrhenius equation was used to quantitatively analyze temperature-dependent developmental tempo. Statistical tests such as the Mann-Whitney U test and bootstrapping were used for hypothesis testing and confidence interval calculation. Image sorting was performed using a custom algorithm based on Euclidean distances and cosine similarities.

The Twin Network accurately predicted the developmental age of zebrafish embryos. Temperature-dependent developmental tempo in zebrafish and medaka followed predictions from classical physical biology theories, with apparent activation energies comparable to other poikilotherms. The method effectively detected natural variability in zebrafish development and identified embryos with abnormal development. Drug-induced phenotypic changes were accurately detected, even without prior knowledge of the specific alterations. The Twin Network automatically generated developmental staging atlases for zebrafish, medaka, three-spined stickleback, and *C. elegans*, demonstrating the generalizability of the approach across diverse species. The accuracy of detecting drug-induced phenotypes depended on both the drug's effect and the number of embryos analyzed.

This study successfully demonstrates the potential of using Twin Networks for the unbiased and automated analysis of developmental processes. The findings highlight the importance of considering the continuous and dynamic nature of embryonic development, rather than relying on discrete stage classifications. The quantitative analysis of temperature dependence provides new insights into the biochemical kinetics underlying developmental tempo. The ability to detect subtle variations and abnormalities opens opportunities for studying developmental robustness, the effects of environmental factors, and the mechanisms of developmental disorders. The *de novo* generation of staging atlases simplifies the study of understudied species and enhances cross-species comparisons. Further improvements to the method could involve generating more generalized models using techniques like Generative Adversarial Networks to overcome the limitation of direct application across different image data types.

The Twin Network approach offers a significant advancement in the field of developmental biology. It provides a standardized, objective, and high-throughput method for analyzing embryonic development. Future work could focus on improving the model's robustness across different species and imaging conditions, exploring integration with other omics data, and applying this method to other complex developmental processes beyond embryogenesis.

The study primarily focuses on zebrafish, with medaka, three-spined stickleback, and *C. elegans* used for demonstrating generalizability. While the results show promise, further validation across a wider range of species is necessary. The direct application of the trained model to different image datasets (species, imaging conditions) is limited, although retraining or fine-tuning could address this.