A team of researchers led by Sarath Sivaprasad from CISPA, along with Hui-Po Wang and Mario Fritz from CISPA and collaborators from HIPS, has developed an AI system capable of automatically detecting developmental abnormalities in zebrafish embryos. The system uses a large-scale, high-resolution image dataset combined with a transformer-based machine learning model to identify signs of toxicity and assess fertility outcomes with high accuracy and efficiency. This breakthrough could greatly accelerate drug screening processes, particularly in early-stage toxicity testing. The research will be presented at the 28th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2025) in Korea. Sivaprasad, whose work focuses on anomaly detection in machine learning, explains that this technique involves training a model to recognize what normal biological development looks like, then identifying any deviations from that norm. Unlike traditional classification that assigns data to predefined categories, anomaly detection focuses on distinguishing between normal and abnormal patterns. In this study, the approach is applied to zebrafish embryos, which are widely used in biomedical research due to their transparent bodies, rapid development, and genetic similarity to humans. Their sensitivity to chemical exposure makes them ideal for high-throughput toxicity screening, a key step in drug discovery. However, current analysis still depends heavily on time-consuming and subjective manual evaluation by experts. The main challenge has been the lack of large, temporally rich datasets suitable for training advanced AI models. To address this, the team at HIPS created one of the most comprehensive image datasets of zebrafish embryonic development, containing over 185,000 microscopic images. Embryos were placed in wells and continuously monitored under a microscope, capturing their development over time. The dataset includes both sequence-level labels for fertility detection and fine-grained temporal annotations for developmental anomalies. Sivaprasad trained a transformer-based neural network on this dataset to analyze both individual images and the progression of development across time. The model achieved 98% accuracy in determining whether an embryo was fertilized and 92% accuracy in detecting developmental abnormalities caused by exposure to the toxic compound 3,4-dichloroaniline. Notably, the AI mimics the way human experts assess developmental changes, enabling early prediction of toxic effects. The researchers aim to scale this approach to screen entire libraries of chemicals, improving both the speed and sensitivity of toxicity testing. The complete dataset will be made publicly available on GitHub, offering a valuable resource for the machine learning and biomedical research communities. It will support benchmarking of new AI methods and help advance more efficient, reliable, and ethical approaches to drug development. The study is published in the conference proceedings under the title "Automated Detection of Abnormalities in Zebrafish Development" (DOI: 10.1007/978-3-032-04981-0_6). The work is part of a broader effort to integrate AI into life sciences, paving the way for faster and more accurate discovery of safe and effective therapies.