Prediction of hERG Liability - Using SVM Classification, Bootstrapping and Jackknifing

Area of Science:

• Computational chemistry
• Drug discovery
• Cardiovascular pharmacology

Background:

• Drug-induced QT prolongation causes life-threatening cardiotoxicity, primarily via blockage of the human ether-à-go-go-related gene (hERG) potassium channels.
• The hERG channel is a critical antitarget in early drug discovery to prevent late-stage development failures.

Purpose of the Study:

• To develop and validate predictive models for hERG channel inhibition using quantitative high-throughput screening (qHTS) data.
• To assess the impact of dataset rebalancing techniques on model performance and identify optimal strategies for predicting hERG liabilities.

Main Methods:

• Screening of 4,323 molecules for hERG channel inhibition using a thallium flux assay in a qHTS format.
• Development of support vector classification (SVC) models to predict hERG channel inhibition.
• Application of Jackknifing and bootstrapping for dataset rebalancing and analysis of their impact on model performance.

Main Results:

• SVC models achieved an averaged area under the receiver operator characteristics curve (AUC-ROC) of 0.93 on the tested compounds.
• Dataset rebalancing techniques did not improve predictive power; optimal cutoffs were found to restore classifier sensitivity and specificity.
• An external validation set of 66 drug molecules showed an AUC-ROC of 0.86 for the SVC model.

Conclusions:

• SVC models are effective tools for predicting hERG channel inhibition early in the drug discovery process.
• Optimal cutoff selection is more critical than dataset rebalancing for binary classifiers predicting hERG liabilities.
• This modeling approach demonstrates significant utility in identifying potential cardiotoxicity risks associated with drug candidates.