Enhancing hERG Risk Assessment with Interpretable Classificatory and Regression Models

Area of Science:

• Computational chemistry
• Pharmacology
• Drug discovery

Background:

• The human Ether-à-go-go-Related Gene (hERG) channel is crucial for cardiac action potential; its inhibition causes potentially fatal cardiotoxicity.
• Current in vitro methods for identifying hERG blockers are costly, necessitating the development of alternative, cost-effective predictive tools.

Purpose of the Study:

• To enhance the Pred-hERG tool for predicting hERG blockage by developing advanced Quantitative Structure-Activity Relationship (QSAR) models.
• To improve early-stage assessment of drug-induced cardiotoxicity through more accurate and interpretable predictive models.

Main Methods:

• Developed new QSAR models incorporating updated data from ChEMBL v30 (14,364 compounds).
• Updated existing binary and multiclassification models and introduced new regression models for predicting hERG activity (pIC50).
• Integrated SHAP (SHapley Additive exPlanations) values for model interpretability and implemented explainable AI (XAI).

Main Results:

• New binary and multiclassification models demonstrated superior performance compared to previous Pred-hERG versions and other public models.
• The optimal regression model achieved an R² of 0.61 and an RMSE of 0.48, outperforming existing literature models.
• Pred-hERG 5.0 provides classification, multiclassification, regression predictions, probability maps, and SHAP value visualizations.

Conclusions:

• Pred-hERG 5.0 offers a user-friendly, reliable platform for early cardiotoxicity assessment via hERG blockage prediction.
• The tool's enhanced predictive accuracy and explainability support informed decision-making in drug development.
• Accessible to users without computational expertise, Pred-hERG 5.0 is freely available for broader research application.