The Catch-22 of Predicting hERG Blockade Using Publicly Accessible Bioactivity Data

Area of Science:

• Computational chemistry
• Pharmacology
• Drug safety

Background:

• Drug-induced inhibition of human ether-à-go-go-related gene (hERG) potassium channels can cause fatal cardiotoxicity.
• Several drugs have been withdrawn from the market due to hERG channel-related safety concerns.

Purpose of the Study:

• To highlight challenges in developing robust classifiers for predicting hERG channel activity using public domain bioactivity data.
• To compare the performance of different machine learning models and data selection strategies for hERG cardiotoxicity prediction.

Main Methods:

• Employed three machine learning methods: nearest neighbors, random forests, and support vector machines.
• Utilized diverse molecular descriptors, activity thresholds, and training set compositions for model development.
• Focused on data filtering criteria and activity threshold settings (binary 1 μM/10 μM vs. single threshold).

Main Results:

• Developed predictive models that demonstrated superior performance in external validations compared to previous studies.
• Found that molecular descriptors had minimal impact on model performance.
• Identified data filtering criteria, activity threshold settings, and structural diversity of blockers as crucial for model robustness.

Conclusions:

• Binary thresholds (1 μM/10 μM) for classifying blockers and nonblockers yield superior predictive model performance.
• Careful selection of training data, particularly the balance of blockers and nonblockers, is critical for reliable hERG activity prediction.
• Public domain data limitations, such as the scarcity of nonblocker data, pose significant challenges for developing accurate hERG classifiers.