hERG toxicity prediction in early drug discovery using extreme gradient boosting and isometric stratified ensemble...

Area of Science:

• Pharmacology and Toxicology
• Computational Chemistry
• Machine Learning in Drug Discovery

Background:

• Blockade of the human Ether-à-go-go Related Gene (hERG) potassium channel by small molecules can cause fatal cardiotoxicity.
• Drug withdrawals due to cardiac side effects highlight the need for early hERG toxicity identification.
• Existing machine learning models face challenges in robustness, class imbalance, and interpretability.

Purpose of the Study:

• To develop a robust and interpretable machine learning model for predicting hERG channel inhibition.
• To improve the identification of potential cardiotoxic compounds in early drug discovery.
• To leverage the largest public hERG inhibition database for enhanced prediction.

Main Methods:

• Integration of eXtreme Gradient Boosting (XGBoost) with Isometric Stratified Ensemble (ISE) mapping (XGB + ISE map).
• Development of an XGBoost consensus model using balanced training sets and diverse variable subsets.
• Application of ISE mapping for applicability domain estimation and prediction confidence evaluation.

Main Results:

• The XGBoost + ISE map model demonstrated robust performance with high sensitivity (0.83) and specificity (0.90).
• ISE mapping improved prediction confidence and compound selection through data stratification.
• Variable importance analysis identified key molecular determinants associated with hERG inhibition (e.g., peoe_VSA8, ESOL).

Conclusions:

• The XGBoost + ISE map strategy offers an effective approach for predicting hERG inhibition.
• This method aids in identifying promising drug candidates with reduced cardiotoxicity risk.
• Enhanced interpretability and robustness address key challenges in current hERG prediction models.