Related Experiment Video
Updated: Feb 26, 2026
10:56
Long-term Behavioral Tracking of Freely Swimming Weakly Electric Fish
Published on: March 6, 2014
13.1K
ADFC-ATP: Attention-Guided Dual-View Fusion and Contrastive Pretraining for Robust Aquatic Toxicity Prediction.
Jixuan Jia1, Xin Yang2, Ying Fang1
1School of Computer Science and Software Engineering, University of Science and Technology Liaoning, Anshan, China.
Journal of Cellular and Molecular Medicine
|February 24, 2026
Summary
A new deep learning framework, ADFC-ATP, improves aquatic toxicity prediction by integrating molecular graph fusion and contrastive learning. This robust and interpretable tool enhances ecological risk assessment for emerging contaminants.
Area of Science:
- Environmental Chemistry
- Computational Toxicology
- Bioinformatics
Background:
- Aquatic ecosystems face threats from chemical pollutants, necessitating better ecological risk assessment methods.
- Current deep learning models for molecular toxicity prediction lack generalizability, interpretability, and robustness, especially with limited data.
- Existing methods struggle to accurately predict the toxicity of emerging aquatic contaminants.
Purpose of the Study:
- To develop a novel deep learning framework, ADFC-ATP, for enhanced aquatic toxicity prediction.
- To improve the generalizability, interpretability, and robustness of molecular toxicity prediction models.
- To provide a computationally efficient tool for assessing the ecological risks of chemical pollutants.
Main Methods:
- Proposed ADFC-ATP framework integrating dual-view molecular graph fusion and contrastive topology learning (NT-Xent loss).
- Utilized structural graph augmentations for robustness and a graph attention encoder for hierarchical substructure pattern learning.
- Employed an adaptive attention-based fusion mechanism combining graph embeddings and fingerprint similarity for toxicity prediction.
Main Results:
- ADFC-ATP demonstrated an average relative improvement of ~10.2% in AUC on four fish toxicity datasets compared to GCN-ST and GCN-MT baselines.
- Ablation studies and attention weight visualization confirmed the importance of scaffold preservation and contrastive regularization.
- The model successfully identified toxicophores, aligning with Quantitative Structure-Activity Relationship (QSAR) principles.
Conclusions:
- ADFC-ATP offers a robust, interpretable, and efficient solution for predicting aquatic toxicity of emerging contaminants.
- The framework serves as a valuable complement to traditional laboratory testing methods for ecological risk assessment.
- The developed model enhances the prediction accuracy and reliability of molecular toxicity assessments in aquatic environments.
