An accurate and interpretable bayesian classification model for prediction of HERG liability

Area of Science:

• Pharmacology
• Computational Chemistry
• Drug Discovery

Background:

• Drug-induced QT prolongation is a serious adverse effect leading to drug market withdrawals.
• Blockade of the hERG potassium channel is the primary cause of drug-induced QT prolongation.
• Early identification of hERG channel blockers is crucial to prevent cardiotoxicity and costly late-stage drug development failures.

Purpose of the Study:

• To develop and validate a predictive model for identifying hERG channel blockers.
• To categorize compounds as active or inactive hERG blockers using a naive Bayes classifier.
• To assess the utility of molecular descriptors in predicting hERG activity.

Main Methods:

• A naive Bayes classifier was constructed using a training set of 1979 compounds.
• The model utilized a universal, generic molecular descriptor system.
• Performance was evaluated using ROC accuracy and validation on an external test set of 66 drugs.

Main Results:

• The naive Bayes classifier achieved an ROC accuracy of 0.87 on the training set.
• The model correctly classified 58 out of 66 drugs in the external validation set.
• Cumulative probabilities provided confidence metrics for hERG blocker predictions.

Conclusions:

• The developed naive Bayes classifier effectively identifies potential hERG channel blockers.
• The combination of atom-typing descriptors and naive Bayes classification allows for model interpretability.
• This approach offers valuable insights for designing safer drug candidates with reduced hERG activity.