Integrating AI and Molecular Modeling for Structural Prediction of a Closed-Pore State of the hERG Channel

The voltage-gated potassium channel hERG (Kv11.1) plays a central role in cardiac repolarisation by mediating the rapid delayed rectifier K+ current (IKr). Blockage of hERG by small molecules can lead to delayed repolarisation, QT interval prolongation, and potentially fatal arrhythmias, making the channel a critical focus in drug safety screening. Despite extensive pharmacological and electrophysiological characterization, a complete structural understanding of hERG gating remains limited by the absence of an experimentally determined closed-state structure. Here, we use AI-based structural modeling to predict and compare candidate closed conformations of hERG. Building on recent work in which AlphaFold2 (AF2) predictions guided by engineered structural templates captured closed and inactivated states, we applied the emerging protein structure predictor, Chai-1, which employs a single-sequence, language model-based approach independent of multiple-sequence alignments (MSAs). The resulting Chai-1 hERG model was compared with the AF2-derived closed structure, a homology model based on the Rattus norvegicus EAG channel, and an experimentally resolved open-state cryo-EM structure. We assessed these models using a combination of all-atom and coarse-grained molecular dynamics simulations, analyzing protein dynamics, pore geometry, gating residue orientation, hydration, and lipid interactions. The Chai-1 and AF2 models displayed strong structural and dynamic agreement, both adopting compact, non-conductive conformations consistent with a physiologically closed-pore state. When run using an MSA, the Chai-1 predictions sample additional conformations with a distinct positioning of the voltage sensing domain (VSD) gating residues. Our molecular dynamics data reveal insights into pore and VSD dynamics, and support a previously identified role for ceramide binding at the M651 residue. Our findings support the plausibility of AI-derived closed-pore state hERG models and underscore the growing potential of deep learning-based protein structure prediction to identify previously uncharacterized, pharmacologically relevant conformations of membrane proteins. Further, our Chai-1 derived closed-pore state model expands our structural insights into hERG gating and may have utility for investigation of drug-hERG interactions.