Machine Learning-Accelerated Analysis of In Utero Embryo Phenotyping in C. elegans for Reproductive Toxicity Assessment.

Predictive new approach methodologies (NAMs) for developmental and reproductive toxicity (DART) assessment are increasingly needed as reliance on conventional mammalian studies decreases and chemical safety evaluation demands continue to expand. Whole-organism NAMs, including Caenorhabditis elegans, provide a scalable non-mammalian strategy because they preserve conserved biological pathways within an intact physiological system. We recently developed vivoDART, a rapid and repeatable C. elegans assay that quantifies in utero embryo development and overcomes key limitations of traditional labor-intensive, multiday C. elegans DART workflows. However, despite its robustness and reproducibility, vivoDART still requires manual analysis of tens of thousands of embryos per chemical, a process that is time-consuming and prone to user-dependent variability. To address this bottleneck, we developed EmbryoMAE-Det, a machine-learning framework trained on ~48,000 manually segmented embryos from 1,547 worms. The model combines self-supervised masked autoencoder pretraining with supervised object detection and classification to identify embryos within the C. elegans uterus and classify them by developmental stage. EmbryoMAE-Det achieved high accuracy (mAP = 88.7%), with AP values of 92.8% and 84.7% for early- and late-stage embryo counts, respectively. Model-derived embryo counts showed low variability with CV%s for technical replicates below 10.4%, sufficient statistical power to detect changes as small as 5-18%, and EC50 values statistically indistinguishable from those obtained by manual scoring. The fully automated workflow reduces analysis time by 1,000×. In summary, this work establishes an integrated whole-organism imaging and machine-learning platform for rapid, reproducible, and high-content DART evaluation using C. elegans as a NAM.