hERG-Att: Self-attention-based deep neural network for predicting hERG blockers

Area of Science:

• Computational Chemistry
• Pharmacology
• Machine Learning

Background:

• The human ether-à-go-go-related gene (hERG) channel is crucial for cardiac action potential regulation.
• Inhibition of hERG channels can lead to cardiotoxicity, making hERG blocker screening essential in drug discovery.
• Conventional screening methods are costly and inefficient, driving the need for advanced in silico models.

Purpose of the Study:

• To develop the first attention-based, interpretable model for predicting hERG blockers.
• To identify critical compound substructures associated with hERG channel blockage.
• To enhance the accuracy and efficiency of hERG blocker screening in early drug discovery.

Main Methods:

• Collected diverse datasets from public and private sources for model training and validation.
• Developed a deep learning model utilizing a self-attention mechanism and Morgan fingerprints for molecular representation.
• Validated the model's performance using rigorous testing and compared it against conventional machine learning approaches.

Main Results:

• The proposed attention-based model demonstrated high performance, accuracy, and F1 scores, exceeding conventional methods.
• The model successfully identified and interpreted important structural patterns within hERG-blocking compounds.
• Validation confirmed the model's optimization and predictive reliability for hERG blockers.

Conclusions:

• The developed model offers a powerful and interpretable solution for predicting hERG blockers.
• This approach can significantly reduce drug discovery costs by enabling accurate and efficient screening.
• The model's interpretability aids in understanding structure-activity relationships for hERG channel interactions.