MTF-hERG: A Multi-Type Features Fusion-Based Framework for Predicting hERG Cardiotoxicity of Compounds.

Area of Science:

• Computational chemistry and toxicology
• Pharmacology and drug discovery
• Artificial intelligence in medicine

Background:

• Human ether-a-go-go-related gene (hERG) channel inhibition causes life-threatening cardiac arrhythmias.
• Accurate prediction of hERG cardiac toxicity is crucial for safe drug development.
• Traditional toxicity assessments are time-consuming and have low throughput.

Purpose of the Study:

• To develop a novel deep learning framework, MTF-hERG, for accurate prediction of hERG cardiac toxicity.
• To enhance drug development efficiency and reduce risks associated with hERG channel blockers.

Main Methods:

• Proposed a multi-type feature fusion framework (MTF-hERG) integrating molecular fingerprints, 2D images, and 3D graphs.
• Employed fully connected neural networks, DenseNet, and Equivariant Graph Neural Networks for feature extraction.
• Utilized deep feature fusion and fully connected layers for classification and regression predictions of hERG activity.

Main Results:

• MTF-hERG achieved high average performance metrics: ACC (0.926), AUC (0.943), AUPR (0.913), RMSE (0.453), and R² (0.681).
• The model significantly outperformed existing state-of-the-art methods on benchmark datasets.
• Visualization revealed key predictive features and decision mechanisms, aiding molecular structure optimization.

Conclusions:

• The MTF-hERG framework demonstrates excellent predictive performance for hERG cardiac toxicity.
• This tool offers robust support for drug development, improving safety and efficiency.
• MTF-hERG has the potential to significantly impact drug discovery and personalized medicine.