Development and Comparison of hERG Blocker Classifiers: Assessment on Different Datasets Yields Markedly Different Results

Area of Science:

• Computational chemistry
• Drug discovery
• Toxicology

Background:

• hERG channel blockade is linked to cardiotoxicity, necessitating predictive models.
• Low throughput of functional assays drives the need for computational approaches.
• Pharmaceutical industry requires models to classify strong (IC50 <1 µM) vs. moderate/weak (IC50 ≥10 µM) hERG inhibitors.

Purpose of the Study:

• To develop and present novel binary classification models for prospective hERG inhibitors.
• To assess the competitiveness of new models against existing literature benchmarks.
• To highlight critical challenges in building and evaluating predictive computational models.

Main Methods:

• Utilized (discretized) 2D molecular descriptors selected via the Winnow algorithm.
• Developed classification models using Random Forest (RF) and Support Vector Machines (SVMs).
• Compared model performance against established methods (Thai & Ecker, Dubus et al.).

Main Results:

• Achieved Matthews Correlation Coefficients (MCC) ranging from 0.40 to 0.87 on external test sets.
• Demonstrated competitive performance compared to existing literature models.
• Observed significant performance variations when training and testing on different datasets, underscoring data dependency.

Conclusions:

• Proposed novel computational approaches for predicting hERG inhibition are competitive.
• Emphasized the importance of rigorous model assessment and awareness of applicability domain limitations.
• Highlighted data variability as a key challenge in developing robust predictive models for drug safety.