AI-driven cardiovascular toxicity assessment of emerging contaminants in water: from deep learning phenotyping to LLM-orchestrated risk evaluation.

Abstract

The widespread occurrence of emerging contaminants (ECs) in aquatic environments is concerning due to their persistence, bioaccumulation, and potential to induce organ toxicity in organisms and humans. Traditional toxicity assessment follows a hierarchical framework, progressing from in vitro cellular assays to in vivo fish and mammalian models. A critical bottleneck across these scales is the conversion of visual observations into quantitative phenotypic data via manual identification and annotation of region-of-interest. Such analyses are time-consuming, labor-intensive, and subjective, thereby limiting large-scale data acquisition. Recent advances in artificial intelligence (AI), particularly deep learning, enable high-throughput, precise phenotypic quantification, transforming image-based toxicity assessment of ECs into standardized and reproducible analysis. This review examines the state-of-the-art deep learning approaches for cardiovascular toxicity assessment across cellular, fish, and murine levels, with a focus on image, fluorescence, and video data analysis. Overall, deep learning models have evolved from low-dimensional analysis and classification toward high-dimensional, multi-parameter phenotyping, integrating single-cell dynamics, organ-level function, and 3D structural reconstruction for precise extraction of toxicity endpoints. Building on these advances, we propose a large language model-driven environmental risk assessment agent that orchestrates analytical workflows progressing from in vitro alert to in vivo validation, integrates multi-tier toxicity data, and generates interpretable and standardized risk outputs. By bridging fragmented experimental tiers and computational analyses, this framework has the potential to shift traditional toxicology from an experience-driven sequential process toward an integrated, knowledge-driven decision-support paradigm, thereby improving the efficiency and scalability of cardiovascular toxicity screening for ECs in water.