Recent Advances in Deep Learning-Based Toxicity Prediction of Emerging Organic Pollutants in Freshwater Ecosystems

Abstract

Purpose of Review

Traditional quantitative structure-activity relationship (QSAR) models struggle to account for the intricate, nonlinear interactions that govern the toxicity prediction of emerging organic pollutants (EOPs) in freshwater ecosystems. This review addresses how deep learning (DL) advances toxicity prediction by leveraging automated feature learning and multimodal integration.

Recent Findings

Recent studies have reported that DL architectures can achieve improved predictive performance compared with traditional models on specific datasets and under specific experimental conditions. Key advancements include graph neural networks (GNNs) directly learning from molecular structures to achieve high predictive accuracy (e.g., R² up to 0.79); multi-modal approaches that integrate chemical, transcriptomic, and environmental data to enhance robustness and ecological relevance; and the application of models to temporal data and omics profiles, offering mechanistic insights into toxic responses.

Summary

DL provides a robust framework for EOP toxicity prediction, significantly improving accuracy and ecological relevance through architectures such as GNNs and ensemble methods. Notwithstanding these advances, challenges persist, such as data heterogeneity, model interpretability, and limited generalizability. To realize the full potential of DL, future efforts must prioritize enhancing data quality, adopting explainable AI (XAI), and improving model applicability.