BPS2026 – Machine learning and interaction features enable prediction of hERG channel blockade

The human ether-à-go-go-related gene (hERG) potassium channel is inhibited by structurally and therapeutically diverse drugs, leading to potentially fatal arrhythmias such as Long QT syndrome. Despite extensive mutagenesis studies highlighting residues like Y652 and F656 as central to binding, the molecular features that distinguish strong from weak blockers remain incompletely understood. Here, we combined long-timescale molecular dynamics (MD) simulations with ensemble docking and machine learning to uncover the interaction principles governing hERG blockade. Clustering of more than 100 ms of MD trajectories using UMAP for dimensionality reduction followed by K-means clustering revealed distinct microstates of the binding pocket, capturing the dynamic hydrophobic cage beneath the selectivity filter that is not apparent from static structures. These microstates formed the basis for ensemble docking of known blockers and non-blockers, demonstrating that dynamic ensembles yield stronger binding affinities than the cryo-EM structure and expose hidden conformations relevant for drug binding. From these docking results, we extracted interaction fingerprints and identified clear patterns that differentiate strong from weak binders. Blockers frequently formed pi-pi and cation-pi interactions with Y652 and F656, along with stabilizing hydrogen bonds with T623, S624, and Y652, which correlated with high binding affinity. To generalize these insights, we trained supervised machine learning models on interaction features derived from ensemble docking. The models successfully distinguished strong from weak binders, with hydrogen bonding, salt bridges, and aromatic interactions emerging as the most predictive descriptors. Together, these results show that integrating MD-derived conformational clustering with interaction-based machine learning provides both predictive accuracy and mechanistic insight, offering a generalizable framework for forecasting hERG channel blockade.