Leveraging Machine and Deep Learning Algorithms for hERG Blocker Prediction

Abstract

The human ether-a-go-go-related (hERG) gene is crucial in enabling the regulation of repolarisation process in the heart. Some chemicals act as hERG blockers, resulting in prolonged QT intervals. Predicting the binding capability of molecules with hERG channels is expected to reduce the burden of cardiotoxicity testing in drug evaluation. The application of machine learning (ML) and deep learning (DL) models in the field of toxicity has gained burgeoning interest. The current study utilises state-of-the-art ML and DL models for predicting the hERG-blocking ability of chemical compounds using a dataset of 8337 molecules. It is noted that spatial relationships within molecules are crucial in predicting hERG blockers. While the threshold for blockers is defined as ≤ 10 µM and for non-blockers, it is 10>μM, our analysis indicates that a threshold of 60- 80 µM provides a more accurate cut-off for non-blockers. This adjustment highlights the importance of concentration levels in reflecting the variability specific to individual interaction sites. The algorithm results show that the internal validation performance of RF, XGBoost, and MLP is strong, with AUC scores of 0.90, 0.90, and 0.87, respectively. In summary, the current study provides a machine learning framework for computation cardiotoxicity assessment by analysis of the hERG blocker concentration cut-offs using different fingerprints at multiple thresholds.