A QSAR model of HERG binding using a large, diverse, and internally consistent training set

Area of Science:

• Pharmacology
• Computational Chemistry
• Drug Discovery

Background:

• The pharmaceutical industry increasingly screens for hERG channel inhibition as part of ADME/Tox profiling.
• hERG channel blockade is a critical factor in drug-induced cardiotoxicity.
• Predictive models are needed to identify potential hERG liabilities early in drug development.

Purpose of the Study:

• To construct quantitative structure-activity relationship (QSAR) models for predicting hERG channel inhibition.
• To systematically evaluate regression models using neural network ensembles.
• To assess the impact of various structure representations and feature selection algorithms on model performance.

Main Methods:

• Utilized a large, diverse dataset of compounds tested in a consistent hERG channel inhibition assay.
• Employed neural network ensembles to capture complex, non-linear structure-activity relationships.
• Applied feature selection algorithms to optimize model performance and prevent overfitting.
• Evaluated models using internal cross-validation and external test sets.

Main Results:

• Neural network ensembles demonstrated strong performance in predicting hERG channel inhibition.
• Feature selection effectively reduced model complexity and improved predictive accuracy.
• The developed models showed good internal cross-validation statistics.
• Models exhibited promising predictivity on unseen external data.

Conclusions:

• A combination of neural network ensembles, feature selection, and aggregation yields robust QSAR models for hERG channel inhibition.
• These models can aid in early identification of potential cardiotoxicity risks.
• The approach offers a valuable tool for pharmaceutical research and development.