Deep learning-based screening of zebrafish embryo malformations for assessing the developmental toxicity and EC50 estimation of disinfection by-products
Year: 2025
Type: article
Abstract: • First to report detailed malformation toxicity of halogenated acetaldehydes in zebrafish. • Constructs a labeled image dataset of zebrafish larvae malformations. • Uses AI to identify organ and morphological deformities in zebrafish larvae. • Automatically fits and outputs EC₅₀ curves based on recognized phenotypes. Chlorine disinfection of drinking water has widely led to the formation of carcinogenic and teratogenic halogenated disinfection by-products, but only a small fraction of them have been identified and undergone systematic toxicological evaluation. In this study, a high-throughput toxicity screening strategy that integrates a computer vision-based model with the zebrafish embryo assay was presented. This framework enables automated identification of teratogenic morphological endpoints from embryo images and facilitates subsequent dose–response analysis for toxicological metrics. Four DBPs (trichloroacetaldehyde, dichloroacetaldehyde, chloroacetaldehyde and tribromoacetaldehyde) were selected for acute exposure. An annotated zebrafish deformity image dataset with literature images and 609 in house images were curated. Channel Attention and Multi-scale Detection module were incorporated into the neural network architecture to improve detection performance. The optimized model demonstrated high classification accuracy across diverse malformation types (mAP50 = 0.873), with strong precision and recall metricsand EC 50 values calculated by the model were highly consistent with those obtained manually. Beyond high-throughput screening, this framework could bridge computer vision with the Adverse Outcome Pathway (AOP) concept, as AI-recognized malformations correspond to key developmental events related to cardiotoxicity and metabolic disruption in the future. This integration enhances the mechanistic interpretability and predictive capacity of AI-based toxicity evaluation, providing a scalable tool for zebrafish-based chemical screening and environmental risk assessment.
Authors: Yizhi Pei, Chunxu Chen, Kongyan Luo, Junmin Gao, Haifeng Zhang
Language: English
Subfield: Cell Biology
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $1,400
OpenAlex ID: W4416866089