BayeshERG: a robust, reliable and interpretable deep learning model for predicting hERG channel blockers

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Unintended inhibition of the human ether-à-go-go-related gene (hERG) ion channel by small molecules causes severe cardiotoxicity, posing a significant challenge in drug development.
• Existing computational models for predicting hERG channel blockage, including deep learning approaches, often lack robustness, reliability, and interpretability.

Purpose of the Study:

• To develop a novel graph-based Bayesian deep learning model, BayeshERG, for accurate and interpretable prediction of hERG channel blockers.
• To enhance the robustness, reliability, and interpretability of hERG channel blocker prediction models.

Main Methods:

• Developed BayeshERG, a graph-based Bayesian deep learning model incorporating transfer learning on a large dataset (300,000 samples) for enhanced predictive performance.
• Implemented a Bayesian neural network with Monte Carlo dropout for uncertainty calibration and utilized global multihead attentive pooling for high-resolution structural interpretability.
• Conducted rigorous internal and external validations, benchmarking against existing publicly available hERG channel blocker prediction models.

Main Results:

• BayeshERG demonstrated superior predictive performance and uncertainty calibration compared to existing models.
• The model's attention mechanism effectively identified essential substructures of hERG channel blockers.
• In vitro experiments validated the model's predictions, confirming its high accuracy and utility.

Conclusions:

• BayeshERG offers a robust, reliable, and interpretable tool for predicting hERG channel blockers, significantly aiding in early-stage drug discovery.
• The model can help mitigate cardiotoxicity risks associated with hERG channel inhibition, thereby maximizing the success rate of new drug development.
• The developed model and its source code are publicly available to facilitate further research and application in pharmaceutical development.