Related Experiment Video
Updated: Jan 17, 2026
05:47
Evidence-based Knowledge Synthesis and Hypothesis Validation: Navigating Biomedical Knowledge Bases via Explainable AI and Agentic Systems
Published on: June 13, 2025
1.3K
hERG-MFFGNN: An Explainable Deep Learning Model for Predicting Cardiotoxicity Using Multi-feature Fusion and Graph
Bingyu Jin1, Jiarun Wang2, Xin Yang3
1School of Electronics and Information Engineering, University of Science and Technology Liaoning, Anshan, 114051, China.
Interdisciplinary Sciences, Computational Life Sciences
|September 22, 2025
Summary
A new deep learning model, hERG-MFFGNN, accurately predicts drug-induced cardiotoxicity by identifying hERG channel blockers. This computational tool aids early drug discovery, reducing risks associated with arrhythmia and torsades de pointes (TdP).
Area of Science:
- Computational chemistry
- Pharmacology
- Cardiovascular drug safety
Background:
- Drug-induced cardiotoxicity, particularly arrhythmia, is a significant hurdle in pharmaceutical development.
- Inhibition of the hERG potassium channel by drug compounds can lead to QT interval prolongation and life-threatening arrhythmias like torsades de pointes (TdP).
- Traditional methods for assessing hERG channel interaction are resource-intensive and unsuitable for high-throughput screening.
Purpose of the Study:
- To develop an accurate and interpretable deep learning framework for predicting hERG channel blockers.
- To enhance prediction accuracy and model generalizability through a multi-feature fusion strategy.
- To provide an efficient computational tool for early-stage identification of potential cardiotoxic compounds.
Main Methods:
- A deep learning framework, hERG-MFFGNN, was developed integrating molecular structural information.
- A multi-feature fusion strategy combined molecular fingerprints, descriptors, and graph neural network-derived topological features.
- An attention mechanism was employed to weight and fuse these features for a comprehensive compound representation.
Main Results:
- hERG-MFFGNN achieved a high predictive performance with an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.909 and an Accuracy (ACC) of 0.854.
- The model demonstrated robust predictive capabilities on both benchmark and external validation datasets.
- The framework provides model interpretability, aiding in understanding the basis of predictions.
Conclusions:
- hERG-MFFGNN serves as an effective computational instrument for the early prediction of hERG channel blockers.
- The developed framework can significantly aid in mitigating cardiotoxicity risks during drug discovery and development.
- The study highlights the potential of integrated deep learning approaches for predicting drug safety profiles.
